YouTube2Text

Subagents vs Agent Teams? 🧠 Hermes Bots, Goal Loops & Kanban Graphs — Transcript

by Wanderloots · 9,470 words · 1,347 segments · language en · Watch on YouTube

Full transcript

  1. 0:00That's just a really powerful way for
  2. 0:01you to significantly reduce the cost.
  3. 0:04There we go. It just dropped from 45,000
  4. 0:06tokens down to 18,000. One agent can
  5. 0:08already delegate to sub agents. So, what
  6. 0:10does making an agent team actually add?
  7. 0:12This is the question I get asked most
  8. 0:14often about multi- aent workflows. Why
  9. 0:16build a team if one capable agent can
  10. 0:18actually get the job done? And honestly,
  11. 0:20sometimes you shouldn't. One powerful
  12. 0:22agent with sub agents can often be
  13. 0:24enough. But other times, a more complex
  14. 0:26workflow can improve the quality and the
  15. 0:28efficiency. Hi, my name is Callum, also
  16. 0:30known as Waterloots, and welcome to
  17. 0:31today's video on agents, loops, and
  18. 0:33graphs. When are multi-agent teams worth
  19. 0:35it? This process also works with agents
  20. 0:37outside of Hermes as well. In today's
  21. 0:39video, we'll explore the different
  22. 0:40levels of agentic workflows. One agent
  23. 0:42with temporary helpers, aka sub agents,
  24. 0:44goal loops that check whether the work
  25. 0:46is actually done, adding specialists
  26. 0:48with their own tools, skills, and
  27. 0:49memories, aka bot mode, and then conbon,
  28. 0:52a way to manage complex agentic tasks.
  29. 0:54Along the way, we'll check what does
  30. 0:56each level actually need and when does
  31. 0:58it make sense for a specific task to add
  32. 1:00more agents or tooling, including a test
  33. 1:02at the end where I give the same task to
  34. 1:04each level and see how it does. Now,
  35. 1:06let's take a look at agents, graphs, and
  36. 1:08loops.
  37. 1:10So, there are three separate questions
  38. 1:11that we're going to ask here. Who does
  39. 1:13the work? How do they check it? And what
  40. 1:15do they do next? So, today I want to go
  41. 1:17through an example scenario that applies
  42. 1:18to almost every type of task working
  43. 1:20with an agent. If information comes in,
  44. 1:23what actions should I take? How should I
  45. 1:25take them? And what does done look like
  46. 1:27for that particular task? For example,
  47. 1:29an email comes in or I take notes on a
  48. 1:31new article and I want to add it to my
  49. 1:33LLM wiki. How does the agent process
  50. 1:35that information in a consistent
  51. 1:37repeatable way so that it actually meets
  52. 1:39the satisfaction that I want for that
  53. 1:41particular task? But first, before we
  54. 1:42get into automations, triggers, task
  55. 1:44graphs, and conbon, let's start with
  56. 1:46just a single agent to see how it works
  57. 1:47and orchestrates with sub aents. will
  58. 1:49build up in complexity until we get to a
  59. 1:51full graph system that lets you
  60. 1:52orchestrate your tasks. If you find this
  61. 1:54video helpful, please like, hype, and
  62. 1:56subscribe as I really appreciate your
  63. 1:57support a lot. If you're looking for
  64. 1:59more ways to support me, please consider
  65. 2:00joining my YouTube or Patreon
  66. 2:01memberships. Members get access to
  67. 2:03insights, tips, and kits for agentic and
  68. 2:05knowledge workflows. I've also put
  69. 2:07together a free article that I'll link
  70. 2:08in the top comment that goes through
  71. 2:10agents, loops, and graphs in a little
  72. 2:11bit more detail for beginners as a
  73. 2:13guide. I'll also include the results of
  74. 2:15the test that I ran so you can go
  75. 2:16through it in a little bit more depth
  76. 2:17and hopefully see how this can help you
  77. 2:19improve the quality of your own
  78. 2:20workflow.
  79. 2:25All right, so the first example that
  80. 2:26we're going to go through is using a
  81. 2:28single agent here, which you're probably
  82. 2:30already familiar with. This would be
  83. 2:32just any type of chatbot, but we're
  84. 2:33going to give this agent the ability to
  85. 2:36have sub agents. So, we can see in the
  86. 2:37bottom left here, we have this option
  87. 2:39for open agents. And we can see that
  88. 2:41right now at the moment we have no live
  89. 2:43sub agents. So when we're working with
  90. 2:44this first example of a single
  91. 2:46specialized agent, we can think of this
  92. 2:47as a parent agent that has the ability
  93. 2:50to create multiple sub aents or children
  94. 2:53agents. So if we go over to capabilities
  95. 2:55on the left here and go over to tools,
  96. 2:57we can search for delegation. And we can
  97. 2:59see that right now task delegation is
  98. 3:01turned off. So I'm going to enable this
  99. 3:03tool right here. And this will enable
  100. 3:05the parent agent to create a series of
  101. 3:08sub aents with specific capabilities. So
  102. 3:10let's get this started.
  103. 3:12It's going to be able to delegate to
  104. 3:14solve our particular problem, which is
  105. 3:16how do I turn incoming information into
  106. 3:18something useful, an action I can take
  107. 3:20or knowledge I can reuse, and what is
  108. 3:22the best strategy for that action
  109. 3:24system? Does it make sense to use a
  110. 3:26single agent with delegation
  111. 3:27capabilities or a multi- aent team with
  112. 3:29different specialized agents? And the
  113. 3:31idea here is that if we can have one
  114. 3:33capable agent that's able to meet my
  115. 3:35standards, that's a complete solution on
  116. 3:37its own. So right now we can see that
  117. 3:39the agent was able to create all of
  118. 3:41these sub aents that are now running in
  119. 3:42parallel. We have three of them. We can
  120. 3:44see on the bottom here that we have
  121. 3:45three different workers active at the
  122. 3:47moment. And what's cool, this is a new
  123. 3:49update. If we click on the three sub
  124. 3:51aents here and click on a specific
  125. 3:52agent, we're able to give instructions
  126. 3:54to this sub agent and steer it. So I
  127. 3:56could get feedback saying, make sure to
  128. 3:59please include citations. And I would be
  129. 4:02able to now steer this agent. And that
  130. 4:05just sent off specific instructions to
  131. 4:07this particular agent that's looking at
  132. 4:09concrete endto-end workflows for inbox
  133. 4:12information. And I could do that for
  134. 4:13each of the different sub agents. And
  135. 4:14now we can see, for example, this
  136. 4:16particular agent, the one that I just
  137. 4:18steered, is now using the research
  138. 4:20grounded citations skill. So I was able
  139. 4:22to give some instruction to the agent as
  140. 4:25it's running to help control the
  141. 4:26direction that it's going in. You can
  142. 4:28click on it here and you can go and see
  143. 4:30how it's thinking. Oh, we just saw this
  144. 4:31one finish right now. Now we have two
  145. 4:33agents instead of three. And we can see
  146. 4:35that one has completed right here and
  147. 4:36the other two are running. So now the
  148. 4:38main agent, the parent agent is waiting
  149. 4:41for the sub agents to complete its work.
  150. 4:43And then it'll synthesize a concrete
  151. 4:44recommendation when they return. Okay,
  152. 4:46here we go. Looks like all three of
  153. 4:47those just finished. So now the main
  154. 4:49agent has gone through and it validated
  155. 4:50a few of the sources and now it is
  156. 4:52creating a report right here. Okay, so
  157. 4:54this gave what looks like a pretty
  158. 4:56comprehensive answer. We can see if I
  159. 4:57scroll all the way to the bottom,
  160. 4:58there's a lot of information here. So
  161. 5:00I'd have to go through that a little bit
  162. 5:01more in depth. And what's nice is if we
  163. 5:02take a look at the bottom here, this
  164. 5:04only used up 36,000 tokens. And the
  165. 5:06reason for that is because the parent
  166. 5:08agent delegated to sub agents and only
  167. 5:11had to give definitions for them to go
  168. 5:13off and run. So the sub agents
  169. 5:14themselves then used up their own amount
  170. 5:16of context. And that's one of the major
  171. 5:18benefits of using a delegation system
  172. 5:20like this is I didn't have to have this
  173. 5:23main parent agent blow through its
  174. 5:24context window. Instead, it could
  175. 5:26delegate to sub aents of lower
  176. 5:28intelligence and then take the
  177. 5:29synthesized reports rather than reading
  178. 5:31all of the pages itself. So, the overall
  179. 5:33idea here is that we're delegating the
  180. 5:35work but not the judgment. And just to
  181. 5:36see how this compares, I'm going to give
  182. 5:38the same prompt asking it to do research
  183. 5:40but do not delegate to sub agent.
  184. 5:43So, the approach where I just used a
  185. 5:44single agent and told it not to delegate
  186. 5:46to sub agents. We can see that used
  187. 5:48102,000 tokens whereas the one that
  188. 5:50delegated to sub agents only used 37,000
  189. 5:53tokens. So keep in mind this is not
  190. 5:55taking into account the sub agent token
  191. 5:57usage. So the overall tokens was larger.
  192. 6:00There was more context used. But that
  193. 6:02context that made it back into the main
  194. 6:04higher intelligence agent was
  195. 6:06significantly compressed relative to the
  196. 6:08agent that just did it all itself and
  197. 6:10had 102,000. If I ran this turn one more
  198. 6:12time, I might have to compress the
  199. 6:14context here and then you lose some
  200. 6:15quality. So that's the first comparison,
  201. 6:17just using a single agent versus using a
  202. 6:19single agent with multiple sub aents. So
  203. 6:21the biggest difference is that one uses
  204. 6:23up a lot more context. But this was also
  205. 6:25a very basic example and the results of
  206. 6:27both of these are basically the same.
  207. 6:29Build an accountable action steward with
  208. 6:31delegation capabilities, not a standing
  209. 6:33team of specialized agents. A permanent
  210. 6:35multi-agent team adds coordination,
  211. 6:36cost, latency, and evaluation problems
  212. 6:38before it has proved its value.
  213. 6:40Multi-agent designs work well when the
  214. 6:42task is genuinely decomposable and
  215. 6:44parallel. And this is a very similar
  216. 6:45answer to what we got with the
  217. 6:47delegation system. Do not begin with a
  218. 6:49peer team of autonomous agents. Delegate
  219. 6:52bounded readonly work to specialists
  220. 6:53when that work is genuinely independent.
  221. 6:55So both of our researchers found
  222. 6:57effectively the same thing that we
  223. 6:58should only delegate specialist work
  224. 7:00when it actually makes sense to do so
  225. 7:02when the complexity needs it. But all of
  226. 7:04this is relative. This is specifically
  227. 7:05for a one-off research task that we just
  228. 7:08did here. As the system starts to scale
  229. 7:09in complexity, we need to start thinking
  230. 7:11about other ways of orchestrating. For
  231. 7:13example, using all the different bots
  232. 7:15like we've set up in previous videos.
  233. 7:17And one thing I want to mention too is
  234. 7:18that when we're delegating to sub
  235. 7:20agents, sub agents know nothing. They
  236. 7:22start with completely fresh
  237. 7:23conversations. So they have zero
  238. 7:25knowledge of the parents conversation
  239. 7:26history, tool calls, or anything
  240. 7:28discussed before they were created for
  241. 7:30the delegation. The sub agents only
  242. 7:32context comes from the goal and the
  243. 7:33context field that the parent agent
  244. 7:35populates when it calls delegate task.
  245. 7:37Delegate task is this tool that we
  246. 7:39enabled at the beginning here. And the
  247. 7:40context is how much information goes
  248. 7:42into all of this. But before we get into
  249. 7:44an actual situation where it makes sense
  250. 7:46to have multi- aent workflows, there's
  251. 7:48one more thing I want to quickly talk
  252. 7:49about, which is a way that we can
  253. 7:51enhance the quality of the output of the
  254. 7:54single agent, whether it's a parent
  255. 7:55agent with delegated sub aents or
  256. 7:57all-in-one. And that actually brings me
  257. 8:00back to the warning of the sub aent
  258. 8:02system, which is that the context comes
  259. 8:04from the goal. So what is a goal?
  260. 8:10So to bring up goal, we can type in
  261. 8:12slash and goal. And this brings up the
  262. 8:14goal command, which is a way that we can
  263. 8:16set a standing goal that Hermes works
  264. 8:18across all turns until it's achieved.
  265. 8:20And basically what that means is before
  266. 8:22we get into the complexity of setting up
  267. 8:24all of these specialist agents and
  268. 8:25figuring out how to coordinate them,
  269. 8:27there's one more step that you should
  270. 8:28try first, which is to enable a goal. A
  271. 8:30goal effectively allows you to link
  272. 8:32multiple prompts together into a bigger
  273. 8:34picture, a broader task than just one
  274. 8:37specific research task. For example,
  275. 8:39instead of managing every step, I hand
  276. 8:41over a bounded outcome. What success
  277. 8:43means, how to verify it, and when to ask
  278. 8:45me. The main agent acts as a judge that
  279. 8:47checks the progress, and the session can
  280. 8:49continue within its own budget. And the
  281. 8:51goal isn't endless research. The
  282. 8:53evidence matters. So before we get into
  283. 8:55setting up a complex system to see if we
  284. 8:58can get more out of our agent, we can
  285. 9:00try using this goal command. So let's
  286. 9:01give our multi- aent a goal here.
  287. 9:05So we can see by using / goal, I now
  288. 9:08have turned this goal active. I've added
  289. 9:10a little bit more detail than just
  290. 9:12saying, hey, can you go research this
  291. 9:13thing? I've said specifically, I want
  292. 9:15different streams of information. Please
  293. 9:17use delegated sub aents of lower
  294. 9:19intelligence to do the research. So each
  295. 9:21one of these should be using at least a
  296. 9:23terra medium instead of terra high. I'll
  297. 9:25talk about model routing in a moment.
  298. 9:26But then I've also said please prepare a
  299. 9:28synthesized report and then once you are
  300. 9:31done that corroborate the report against
  301. 9:32practical workflows from community best
  302. 9:34practices in Hermes. You also must make
  303. 9:37sure the report is less than 500 words
  304. 9:38and has clear citations. We can see that
  305. 9:40the sub agents are all running right
  306. 9:41now. So what I've effectively done here
  307. 9:43is I've told the agent what done looks
  308. 9:46like. So it's taken all of these
  309. 9:48criteria and it will just keep running.
  310. 9:50We can see turn one of 20. The default
  311. 9:52limit is 20 turns. The agent is able to
  312. 9:54loop itself 20 times, checking to make
  313. 9:57sure that the information and the
  314. 9:58research and the output meet the quality
  315. 10:01standard that I've requested. So, this
  316. 10:02is a really great way for you to
  317. 10:04increase the quality of your work
  318. 10:05without necessarily introducing a lot of
  319. 10:07complexity. I honestly use this all the
  320. 10:09time, especially for coding. So, I can
  321. 10:11also add criteria on here. For example,
  322. 10:13we must have at least five peer-reviewed
  323. 10:16sources. And I can add that information
  324. 10:18to the criteria as the agent continues
  325. 10:21running. And that now becomes a new
  326. 10:23definition of what done looks like. So
  327. 10:25we can see here it took my criteria into
  328. 10:28account because now the criteria is we
  329. 10:30must have at least five peer-reviewed
  330. 10:31sources. And if we look here, it only
  331. 10:33produced eight sources and it doesn't
  332. 10:35look like they're peer- reviewviewed. So
  333. 10:36now the agent has started up its next
  334. 10:38task on turn two and turn three and it's
  335. 10:41delegated to another series of sub aents
  336. 10:43to now cross-reference the output
  337. 10:45against peer-reviewed academic sources.
  338. 10:48So it says it's now added an independent
  339. 10:50organizational coordination and team
  340. 10:51cognition literature to the evidence set
  341. 10:53and it's awaiting the peer review before
  342. 10:55it will work again. As this is working,
  343. 10:57I can think of more criteria and I can
  344. 10:58add this here and the agent will just
  345. 11:00keep looping. The more specific you are
  346. 11:02on what done looks like at the
  347. 11:04beginning, what the actual goal is, the
  348. 11:06better this process will work. Depending
  349. 11:07on the type of task, depending on the
  350. 11:09type of budget you have, depending on
  351. 11:10your goals, you're going to have
  352. 11:12different ways of doing this. So, it's
  353. 11:13worth experimenting to see what actually
  354. 11:16makes the most sense for your particular
  355. 11:17situation. So, it's kind of a handsoff
  356. 11:19method of operating. If you've
  357. 11:21established a good enough goal at the
  358. 11:23beginning, the agent can run on its own.
  359. 11:24In some systems, I've had my coding
  360. 11:26agents run for five or six hours
  361. 11:28straight, working on a coding system,
  362. 11:30delegating to lower intelligence sub
  363. 11:32aents, and it keeps automatically
  364. 11:34testing the code to make sure it works
  365. 11:36before it presents it back to me for
  366. 11:37approval. And that brings me to an
  367. 11:39important point because if you're
  368. 11:41running a goal here and it runs for
  369. 11:42hours and you have this set to Astra for
  370. 11:45example or to Fable or whatever your
  371. 11:47Frontier model is, it's possible that a
  372. 11:49goal could just completely blow through
  373. 11:51your quota in a very short period of
  374. 11:53time, especially if you're not sitting
  375. 11:55there monitoring it. So this brings me
  376. 11:56to an important question on who should
  377. 11:58actually be doing the work here.
  378. 12:01So what we can see is the sub agents
  379. 12:03here. It just says GPT 5.6 Terra. So I
  380. 12:06can't actually see at the moment if it
  381. 12:07is medium or lower. So as part of the
  382. 12:09goal here, I could have specified rather
  383. 12:11than just please use delegated sub aents
  384. 12:13of lower intelligence. I could have said
  385. 12:15please use a local model or please use
  386. 12:18Luna only for the research and then you
  387. 12:20as the higher intelligence model can
  388. 12:22synthesize it. So this is where you can
  389. 12:24start being more specific with your
  390. 12:25prompting and specific with your
  391. 12:27orchestrating through a single agent
  392. 12:29delegating to lower intelligence sub
  393. 12:31agents to significantly preserve your
  394. 12:32quota. makes a massive difference
  395. 12:34because keeping an agent working isn't
  396. 12:36the same thing as using its
  397. 12:38intelligence. Well, I want the
  398. 12:39orchestrator agent, the main agent to
  399. 12:42make the judgments, not to carry every
  400. 12:44source detail or do every little action
  401. 12:46itself. So, we can use these lightweight
  402. 12:48helpers for bounded extractions with
  403. 12:51limited actions. And what I mean by that
  404. 12:53is we can be specific about what type of
  405. 12:55model is used in which type of
  406. 12:57situation. So, this can apply to the
  407. 12:58original goal that I just gave here. I
  408. 13:00can be specific like I mentioned of what
  409. 13:02type of model I want to be used. I can
  410. 13:04bring in the agents.mmd file where for
  411. 13:07this particular profile this particular
  412. 13:09bot or Hermes agent I can specify inside
  413. 13:12of the agents.mmd or inside of for
  414. 13:14example the soul.MD I can say when using
  415. 13:18sub agent delegation always use lower
  416. 13:21intelligence models. By including it
  417. 13:23here, it means that every single time
  418. 13:25this particular bot, this particular
  419. 13:27agent wants to run a task that involves
  420. 13:30some sub aent delegation, it will always
  421. 13:32know to use lower intelligence models.
  422. 13:34And I can go even a step further by
  423. 13:36going under settings and model and I can
  424. 13:38choose the auxiliary model for different
  425. 13:40types of tasks where for example maybe
  426. 13:42for image analysis I don't want to use
  427. 13:44my main model. I can change this and use
  428. 13:46a local model or I could use a cheaper
  429. 13:48model. So you can go in and you can
  430. 13:50start limiting the specific actions to
  431. 13:52specific models to help conserve the
  432. 13:54quota and intelligence of the main agent
  433. 13:57for when it actually needs it.
  434. 14:00And honestly, this did a really good
  435. 14:01job. It was able to specifically limit
  436. 14:03the report to just a couple hundred
  437. 14:05words instead of the original thousands
  438. 14:07and thousands of words it gave me off of
  439. 14:09the first one. So it wasn't just about
  440. 14:11adding that limitation at the beginning.
  441. 14:12It was framing the goal that I wanted it
  442. 14:15to go in a particular order for cross
  443. 14:17referencing and corroborating the report
  444. 14:19and then make sure that it was 500
  445. 14:21words. But honestly, this example here
  446. 14:23is kind of disconnected from my actual
  447. 14:26workflow. This might be the generic best
  448. 14:29answer of when delegation beats a
  449. 14:31specialist team where we want to have an
  450. 14:32orchestration and delegation for a
  451. 14:35single evolving state, sequential or
  452. 14:37predictable work. And in contrast, you
  453. 14:38might want a multi- aent team for
  454. 14:40breadth first research, different types
  455. 14:42of due diligence, pipelines with
  456. 14:44different specialist identities. So this
  457. 14:46all makes sense, but everything I've
  458. 14:47just said here is not very specific to
  459. 14:49me. It doesn't know who I am or what I
  460. 14:51want. And that brings me back to Hermes
  461. 14:54corroboration where delegate task brings
  462. 14:56up a fresh context fork join primitive
  463. 14:59whose children return summaries. So, it
  464. 15:01fits bounded research, review, and
  465. 15:03repair work, but it doesn't necessarily
  466. 15:05make sense for other types of situations
  467. 15:07that require more specialization and
  468. 15:10more context and more knowledge because
  469. 15:12like I mentioned, sub agents have zero
  470. 15:14knowledge from the parents conversation
  471. 15:15history. They only have what comes from
  472. 15:17the goal and the context that the parent
  473. 15:19gives them. But what if we start getting
  474. 15:21into more specific context or more
  475. 15:23custom situations that involve other
  476. 15:25types of context outside of what the
  477. 15:27parent even knows? We can see here the
  478. 15:28decision test that the answer says is
  479. 15:30add a specialist if it is independent
  480. 15:32needs different context tools and policy
  481. 15:34and beats the single agent baseline on
  482. 15:36quality coverage latency and total
  483. 15:38review cost. So in other words the
  484. 15:40answer is that we should add specialists
  485. 15:42if we want it to keep its own working
  486. 15:44knowledge and have access to different
  487. 15:45context tools and policies perhaps have
  488. 15:48a different memory system rather than
  489. 15:50being spawned as a brand new agent every
  490. 15:52single time. And I note that the Hermes
  491. 15:54corroboration here mentions Kambban or
  492. 15:56Conbon depending on how you want to
  493. 15:58pronounce it. But before we get into
  494. 15:59conbon, I thought that we could take a
  495. 16:01quick look at Hermes bot mode actually
  496. 16:03sets you up for success when you start
  497. 16:05using conbon because it helps you
  498. 16:07prepare the specialist profiles. Like we
  499. 16:09can see down here I have orchestrator,
  500. 16:11researcher, and librarian. So let's take
  501. 16:12a quick look at bot mode.
  502. 16:18So now we can go from a single agent
  503. 16:20operating in a loop with delegated sub
  504. 16:22aents to a graph. That's where profiles
  505. 16:24come in. Okay. So this is bot mode. So
  506. 16:26we can see here a bot is just a profile.
  507. 16:29It's the same list of profiles that I
  508. 16:30have along the bottom here. A bot or a
  509. 16:33profile can retain its own skills,
  510. 16:35memory, and working context with a model
  511. 16:37suited to its role. Each of these
  512. 16:39different bots here can have a default
  513. 16:41model that's different. For example,
  514. 16:43maybe my orchestrator has one of the
  515. 16:44most powerful models available, a
  516. 16:46frontier model, but maybe the researcher
  517. 16:48needs a much less powerful model. And
  518. 16:50then I can use a feature like the group
  519. 16:51chat to coordinate how they work
  520. 16:53together. So rather than a sub agent
  521. 16:55starting with zero knowledge, we can
  522. 16:57configure what type of knowledge and
  523. 16:58skills and memory we want available for
  524. 17:00each specialist. And that knowledge and
  525. 17:02all of that context and capabilities
  526. 17:04survives the chat conversation. It
  527. 17:07maintains it over time. And since Hermes
  528. 17:09is a self-learning system, it can also
  529. 17:11learn more skills as well, getting
  530. 17:13better at its particular specialized
  531. 17:14task. And we can take a quick look at a
  532. 17:16group chat here where we can see, for
  533. 17:17example, I have the librarian, the
  534. 17:19orchestrator, and the researcher, and
  535. 17:20they're all working together to solve
  536. 17:22the original task that I put out. So,
  537. 17:24you can think of this as a communication
  538. 17:25graph where we've set up a group chat
  539. 17:27here, and the group chat has particular
  540. 17:29members, and they work together,
  541. 17:30communicating back and forth with one
  542. 17:32another to solve a particular problem.
  543. 17:34So, this is where we go from not just
  544. 17:35having a loop, but going from a loop to
  545. 17:38a graph. Now, I've already talked a lot
  546. 17:39about this in the last two videos. In
  547. 17:41the first video, I gave an overview on
  548. 17:43what bot mode is, how to get it set up,
  549. 17:44and how to customize each of these
  550. 17:46individual profiles. So, I'm not going
  551. 17:48to touch on that today. And then in the
  552. 17:50second video, we did a specific research
  553. 17:52upgrade where I updated the capabilities
  554. 17:53of the researcher to be a much more
  555. 17:55powerful tool. The main skill that I
  556. 17:56gave the researcher as an upgrade
  557. 17:58allowed it to check both academic
  558. 17:59sources and state-of-the-art sources and
  559. 18:01then corroborate the two of them. I just
  560. 18:03wanted to quickly show you this because
  561. 18:04it is a form of graph in the sense that
  562. 18:06the orchestrator is able to communicate
  563. 18:08and message back and forth with the
  564. 18:10different sub aents, but it's not the
  565. 18:12same as a full task graph like we can
  566. 18:14get into with Conbon. All you need to
  567. 18:16know at this point is that the
  568. 18:17orchestrator is the one that I message.
  569. 18:19It's my human proxy, the one that
  570. 18:20controls the other agents. The
  571. 18:22researcher is able to investigate. So,
  572. 18:24in the example of information coming in,
  573. 18:26the orchestrator would create the task
  574. 18:28and then give it to the researcher to go
  575. 18:30investigate. And then the librarian
  576. 18:31works with my existing LLM wiki to take
  577. 18:34the output research that I've approved
  578. 18:35as a human and file it into my LLM wiki
  579. 18:38or my knowledge base in Obsidian. So in
  580. 18:40a graph, each node can be a single agent
  581. 18:42or action or a bundle of actions. For
  582. 18:45example, like an agent loop built into
  583. 18:47the group as part of a bigger macro
  584. 18:49loop. So in group chats, one of the key
  585. 18:51limitations is that they only have the
  586. 18:53ability to go back and forth three times
  587. 18:55before they have to stop and check what
  588. 18:57you actually want.
  589. 18:59So before we actually get into running
  590. 19:01these full control graphs, I just wanted
  591. 19:02to quickly talk about a couple examples
  592. 19:04on how this fits together so that you
  593. 19:06can see what the graph might look like
  594. 19:08before we get into how to set it up just
  595. 19:10so it makes a little bit more practical
  596. 19:11sense. So imagine you have an email come
  597. 19:14in and you want to have a decision.
  598. 19:16Maybe you want to have a draft created,
  599. 19:17maybe you want to have some research
  600. 19:19done, maybe both. So the email can come
  601. 19:21in and it can trigger an agent like the
  602. 19:23researcher agent to pull in the relevant
  603. 19:25context and the evidence by doing some
  604. 19:27research. Once it's done, it can then
  605. 19:29create a brief like a research brief and
  606. 19:31draft a reply and then hand it off to
  607. 19:33you for review. If you don't think it's
  608. 19:34good enough quality or you think it's
  609. 19:36missing something, you can provide some
  610. 19:37more context, some more commentary, or
  611. 19:39steer it, and you can have it go back
  612. 19:41and loop again. So this brings in a
  613. 19:43loop, but it's a human in the loop. Then
  614. 19:44if it meets your approval, you can have
  615. 19:46it create the draft or even potentially
  616. 19:48send it right away if you've already
  617. 19:50reviewed the final result. It's up to
  618. 19:52you if you want to have the send
  619. 19:54capability given to the agent or if you
  620. 19:55want to click the button. As another
  621. 19:57example, let's say you've come across a
  622. 19:59great article or a YouTube video or some
  623. 20:02other resource and you've taken some
  624. 20:04notes on it. You can drop in the raw
  625. 20:05source and your notes and you can send
  626. 20:07it off to the librarian. The librarian
  627. 20:09agent can try and find the fit. How does
  628. 20:11this fit into your existing knowledge
  629. 20:13base? It'll check the context
  630. 20:14connections and draft an entry for you
  631. 20:17for your review to see is this good
  632. 20:19enough for it to make it into your
  633. 20:21knowledge base into your curated
  634. 20:23library. If it's not good enough, you
  635. 20:25can suggest revisions and the librarian
  636. 20:27will update the draft entry and then
  637. 20:28pass it back to you for review. Once you
  638. 20:30approve it, the librarian can then file
  639. 20:32it into the library and connect it to
  640. 20:34the rest of your research. And as a more
  641. 20:35complex system, we can actually start
  642. 20:37linking these loops together. For
  643. 20:38example, you can propose a question or a
  644. 20:40topic to the researcher. The researcher
  645. 20:42can go through a research loop checking
  646. 20:44sources, checking it against the quality
  647. 20:46that you've given. This could be a goal
  648. 20:48loop. And then when it's done all of its
  649. 20:49research, it can pass it off to you as
  650. 20:51the human to validate that the
  651. 20:54information is good enough. You can add
  652. 20:55your own notes, connect it to that cited
  653. 20:57source and report and then hand it off
  654. 20:59to the librarian. The librarian can then
  655. 21:01see how that fits into your existing
  656. 21:03knowledge base and prepare a draft for
  657. 21:05your review on whether or not it should
  658. 21:06make it into the library. Again, you can
  659. 21:08go through a curation loop, but this one
  660. 21:10has a human in the loop as part of the
  661. 21:12core loop itself. Once you've approved
  662. 21:14it, the librarian can then go and file
  663. 21:15this into the library. And then, as a
  664. 21:17coding example, let's say you're
  665. 21:18building an app and you want to add the
  666. 21:20ability to export to Markdown. You can
  667. 21:22hand this off to your orchestrator and
  668. 21:24explain what you're looking for. And the
  669. 21:25orchestrator can understand it and then
  670. 21:27delegate it to the implement. The
  671. 21:29implementer can then go through a goal
  672. 21:30loop where it builds it, checks to see
  673. 21:33if it works, and if it doesn't, then it
  674. 21:34repairs it and then it checks it again.
  675. 21:36So it can loop as many times as you want
  676. 21:38until the task is actually done. Once
  677. 21:40it's done, you can pass it off to a
  678. 21:41reviewer for an independent review. Kind
  679. 21:43of like a second version of the
  680. 21:44implementer, but specifically for
  681. 21:46quality assurance, who can then pass it
  682. 21:47off to the orchestrator and check to
  683. 21:49make sure it matches the original goal.
  684. 21:51If it does, then it can hand it off to
  685. 21:52you for either a pull request or for you
  686. 21:54to do something with that particular
  687. 21:56output. In the original research, the
  688. 21:58Hermes cooperation actually surfaced
  689. 22:00that conbon is Hermes durable
  690. 22:02alternative with name profiles, shared
  691. 22:04state, dependencies, retries, and work
  692. 22:06that survives the lead session. So, what
  693. 22:08that means is it's a lot easier to start
  694. 22:10and stop and add more tasks and modify
  695. 22:12and just it gives you a lot more control
  696. 22:15on the actual task system that's
  697. 22:17happening here and the way the different
  698. 22:18agents fit into it. That's why it's
  699. 22:20called a task graph. But you might be
  700. 22:21wondering, isn't that kind of similar to
  701. 22:22what I already showed you where we have
  702. 22:24the orchestrator or the main agent
  703. 22:26delegating to sub agents?
  704. 22:31So the Hermes docs has a good breakdown
  705. 22:33here of what the difference is because
  706. 22:35they are actually quite different
  707. 22:37systems. In delegate task, the parent is
  708. 22:39blocked from doing things until the sub
  709. 22:41agent actually returns. Whereas in
  710. 22:43Conbon, the orchestrator can just set
  711. 22:45the agent off, can assign the task, and
  712. 22:47it can just go and do its own thing, and
  713. 22:49you don't have to have anything make it
  714. 22:50back to the orchestrator unless you want
  715. 22:52it to. Rather than having some random
  716. 22:54anonymous sub agent that starts from
  717. 22:56scratch, you have a named profile with
  718. 22:58persistent memory. So that's a pretty
  719. 22:59important element. Also, for
  720. 23:01resumability, if the sub agent fails to
  721. 23:03do its task, it's just done. It just
  722. 23:05comes back and says incomplete. Whereas
  723. 23:07in conbon, you can set different types
  724. 23:09of restoration processes. So it can be
  725. 23:12blocked and unblocked and then rerun
  726. 23:13again from where it left off. So this is
  727. 23:15what I mean when I say it's more
  728. 23:17durable. It has the ability to survive
  729. 23:19errors and keep going. With sub agents,
  730. 23:21you can steer, you can add some
  731. 23:23feedback, but it's not necessarily
  732. 23:25actually adding a human in the loop. But
  733. 23:27in conbon, we can have specific actions
  734. 23:29or task actually block from happening
  735. 23:32until the human comes in and says go.
  736. 23:34And with delegation, the audit trail,
  737. 23:36all of the context can be lost when the
  738. 23:39system actually compresses. This is what
  739. 23:41I was talking about with the context
  740. 23:42window on the bottom here. If this gets
  741. 23:43compressed, then if there was a sub
  742. 23:45agent within that compression, you can
  743. 23:47lose the context associated with it.
  744. 23:49Whereas in conbon, it's stored in a
  745. 23:51SQLite database. So you get access to
  746. 23:53that forever. You can go back and take a
  747. 23:54look. So a good summary here is that
  748. 23:56delegate task is a tool call. It's just
  749. 23:58one function. Whereas conbon is a work Q
  750. 24:01where every handoff is a row in a
  751. 24:02database that any profile or human can
  752. 24:05see and edit. So there's a lot more
  753. 24:06transparency that happens here, which is
  754. 24:08why it starts to get into being a proper
  755. 24:10control graph or task graph. Thankfully,
  756. 24:12Hermes has a built-in feature called
  757. 24:13Conbon. So let's take a look at that.
  758. 24:17So the first thing we need to do for
  759. 24:18conbon is you may not have this
  760. 24:20appearing on the side here. So we can go
  761. 24:21up to capabilities and then go over to
  762. 24:23plugins. And we can see that we have
  763. 24:24three built-in plugins that come with
  764. 24:26Hermes. The first is bot mode that comes
  765. 24:28turned on like we've already looked at
  766. 24:30and the second is the conbon. So make
  767. 24:31sure this is turned on here. And then
  768. 24:33you should see this appear on the lefth
  769. 24:35hand side. So if we go over to conbon
  770. 24:36here, we are not going to be the ones
  771. 24:38running things. The agent is. All we
  772. 24:40need to do is put a card into the ready
  773. 24:42section, assign it to a particular
  774. 24:43agent, and the agent will pick it up in
  775. 24:45a minute. And there's even a triage
  776. 24:47section. So this is where if you're not
  777. 24:49quite sure exactly what the task is, you
  778. 24:51can have an agent rewrite the idea into
  779. 24:53a proper task first. So that's actually
  780. 24:55a really important part of it, and I'll
  781. 24:57show you how we can change the settings
  782. 24:58for that in a moment. So triage is for
  783. 25:00raw ideas where that the specifier is
  784. 25:02going to flush out the spec and then it
  785. 25:04will move it through the system. To-do
  786. 25:06means it's waiting on dependencies or
  787. 25:08unassigned. So we can create task and
  788. 25:10not assign it and that means the
  789. 25:11dispatcher won't send it off to go run.
  790. 25:14We can schedule which is we want a task
  791. 25:16to run at a particular time. So it could
  792. 25:18be for example check the latest series
  793. 25:20of emails and see if something has come
  794. 25:23in that I've been waiting for. If so
  795. 25:24then execute the next task. Ready means
  796. 25:27that the dependencies are satisfied.
  797. 25:28it's been assigned to a profile and the
  798. 25:30dispatcher runs it. Running means it's
  799. 25:31claimed by a worker and an agent's
  800. 25:33actually working on it. Block means it's
  801. 25:34waiting for human input. So, this brings
  802. 25:36that human in the loop. And then review
  803. 25:38means that there's a review agent
  804. 25:39specifically checking out the work. So,
  805. 25:41you can establish which agent you want
  806. 25:42to operate as the reviewer. And this is
  807. 25:44the completed section. But with all that
  808. 25:46said, why don't we run a quick test
  809. 25:47task?
  810. 25:50So, I'm going to leave auto decompose
  811. 25:51triage task turned on. And I'm going to
  812. 25:53click new task. I'm just going to call
  813. 25:54this test task. I'm going to leave this
  814. 25:56assigned to the Wanderloots tutorials
  815. 25:59assigne and click create task. So what
  816. 26:01we should see now is we've added this
  817. 26:04card to triage and the dispatcher is
  818. 26:06going to recognize this and hand it over
  819. 26:08to the auto decomposer within a minute.
  820. 26:10And I'll explain all of that in a
  821. 26:11moment. So triage is a place for raw
  822. 26:13ideas where the specifier will flesh it
  823. 26:15out. So we just saw that it went from
  824. 26:16this raw idea into running. But this
  825. 26:19task here is clarify the scope because
  826. 26:21it sounds like a test. So that's where
  827. 26:23putting something in triage will
  828. 26:24automatically assign it to the default
  829. 26:26that we have set up and it will run over
  830. 26:28here. So this should get blocked because
  831. 26:30this was a task that had no clear
  832. 26:32outcome. So this is where the human in
  833. 26:34the loop starts to come in because it
  834. 26:36basically said there's no body. There's
  835. 26:37no information here. We don't know what
  836. 26:39they want.
  837. 26:41So the dispatcher basically runs inside
  838. 26:43of their Hermes gateway inside of the
  839. 26:45back end. And you don't have to install
  840. 26:47anything. It just checks every 60cond
  841. 26:49has there been a new card created? If
  842. 26:51so, let's move through the conbon loop.
  843. 26:53Let's go through the graph. The idea is
  844. 26:55that once this task gets created as a
  845. 26:57card, the dispatcher will see, oh,
  846. 26:59there's a new card. Let me run the
  847. 27:00system for you. But if we go back to my
  848. 27:02task for a second, we can see here that
  849. 27:04we have a rough idea. The specifier will
  850. 27:06flesh it out. So, we can also see that
  851. 27:08we have auto decomposed the triage
  852. 27:10tasks. So, this is where if we go over
  853. 27:12to settings again, and inside of models,
  854. 27:14we go down to auxiliary models. We can
  855. 27:17see here that we have two different
  856. 27:19models set, two different auxiliary
  857. 27:20models that we can control here. The
  858. 27:22first is the triage specifier and this
  859. 27:24is the one that will take the idea, the
  860. 27:26goal that you have and frame it as a
  861. 27:28better task, which will then get passed
  862. 27:29on to the conbon decomposer to break
  863. 27:32that task into a task graph. So you can
  864. 27:35control what you want these models to be
  865. 27:36here. Perhaps the specifier is a very
  866. 27:38lightweight model because you're
  867. 27:39basically just using text to explain
  868. 27:42text a little bit better. But for the
  869. 27:43decomposer, you probably want a much
  870. 27:45more powerful model. And Hermes docs
  871. 27:48actually has a strategy for this where
  872. 27:49they say decomposing a project, breaking
  873. 27:52it into well scoped cards takes frontier
  874. 27:54level judgment using the most
  875. 27:56intelligence that you have. But
  876. 27:58executing a card that has a clear goal,
  877. 28:00context, and handoff doesn't really need
  878. 28:02that same level of intelligence,
  879. 28:03especially if it's running in a loop.
  880. 28:05Workers are where the vast majority of
  881. 28:06tokens are spent. So that's typically
  882. 28:08where the cost is. So that's what I'm
  883. 28:10getting at when I talk about this
  884. 28:11decomposer here. We can set this model
  885. 28:13to whatever we want. I could set this
  886. 28:15for example to GPT6 Astra if I wanted
  887. 28:17to. And what would happen is now the
  888. 28:19conbon decomposer, this decomposer
  889. 28:21setting that we have here would be using
  890. 28:23Astra to convert whatever task I put
  891. 28:26inside of this task card into a series
  892. 28:28of subtasks that then get assigned to
  893. 28:30the worker profile, which is wonder
  894. 28:32tutorials that I currently have set at
  895. 28:34Terra. So that's just a really powerful
  896. 28:36way for you to significantly reduce the
  897. 28:39cost of your operations by having a
  898. 28:42powerful planning model and a less
  899. 28:44powerful execution model or implement or
  900. 28:47worker agent.
  901. 28:49And we can see up at the top here that
  902. 28:51we have a default board. So what we can
  903. 28:53do is we can click on the default. We
  904. 28:54can rename that board if we want to or
  905. 28:56create a new one. So I can for example
  906. 28:58create a new board called test create
  907. 29:00board. And now when I click on this I
  908. 29:01can switch between different boards. And
  909. 29:04what's cool too is within that test
  910. 29:06board, you can assign it to a specific
  911. 29:07project. So what that means is you can
  912. 29:09have different tasks and different
  913. 29:10graphs and different configurations and
  914. 29:12settings and operations that you want
  915. 29:14for particular boards and all of that
  916. 29:16can be controlled at the project level.
  917. 29:18So you can create a new project like for
  918. 29:20example maybe one is your admin project
  919. 29:22and it deals with everything related to
  920. 29:23email. Maybe another one is a research
  921. 29:25project and it deals with all of your
  922. 29:26research that you're bringing in. Maybe
  923. 29:28another is for a particular coding
  924. 29:30project or app that you're working on
  925. 29:31and all of that can maintain the context
  926. 29:33within this particular project and then
  927. 29:35you can assign a particular board
  928. 29:37associated with that project to run
  929. 29:39these tasks.
  930. 29:41So we can open up the orchestration
  931. 29:43settings here and what that does is it
  932. 29:44brings up a list of all of the profiles.
  933. 29:46It brings up all of the different bots
  934. 29:48or agents that we have available with
  935. 29:50all of the particular settings and
  936. 29:52skills and memory and context that we
  937. 29:55were talking about earlier. Each of
  938. 29:56these operates as a different
  939. 29:58specialized agent. So there's a few key
  940. 30:01settings we need to take a look at. The
  941. 30:02first is the orchestrator profile. So
  942. 30:04right now it's defaulting to default.
  943. 30:06You can change which agent you want to
  944. 30:08operate as your orchestrator. This is
  945. 30:10the orchestrating agent that will decide
  946. 30:12which agent gets which task. And those
  947. 30:14tasks get assigned to the assigne which
  948. 30:16you can set which one you want to be the
  949. 30:18default. Maybe I want to create an
  950. 30:19implement profile whose sole job is to
  951. 30:21execute on things. in a similar way that
  952. 30:23the researcher is the one that
  953. 30:24researches things and the librarian is
  954. 30:26the one that files things in my LLM
  955. 30:28wiki.
  956. 30:33So for this one I'm going to leave on
  957. 30:35auto decompose just so you can see how
  958. 30:36this works. I'm going to click new task.
  959. 30:38I'm going to call it multi- aent
  960. 30:40research and I'm going to include the
  961. 30:41description of the same prompt that I
  962. 30:43gave previously. So here we can set the
  963. 30:45priority. This basically just controls
  964. 30:46the order if you have multiple tasks
  965. 30:48going in there at once. And then we can
  966. 30:49have the workspace that we want it to
  967. 30:51run in. I'm going to leave the current
  968. 30:53assigne as the waterlude tutorials as
  969. 30:55the default even though I'm asking for
  970. 30:57research. So this is where the task
  971. 30:59decomposer should be able to recognize
  972. 31:02that this is actually a research task,
  973. 31:03not an implementation task for the
  974. 31:05tutorials agent. I can list specific
  975. 31:07skills if I want. I can change the model
  976. 31:09that I want this to run at or just have
  977. 31:11it run at the profile default. And
  978. 31:13importantly, I can have this run on goal
  979. 31:15mode. So by default, each worker agent
  980. 31:17gets one shot at its card to do the
  981. 31:19work. But turning goal mode into true
  982. 31:21for that particular task lets that
  983. 31:23worker run in a goal loop like I've
  984. 31:24already shown you. And this is its own
  985. 31:26self-contained goal loop. It doesn't
  986. 31:28connect to the existing goal that we
  987. 31:29have elsewhere. And what's cool too is
  988. 31:31we can also estimate the cost of what
  989. 31:33this task is going to be. So I can click
  990. 31:34estimate and then it gives me a list of
  991. 31:36the tokens that it thinks it's going to
  992. 31:37cost for this particular prompt here. So
  993. 31:39it took this task and it analyzed it to
  994. 31:42give you an estimate. So it says it's
  995. 31:43going to be a broad ambiguous research
  996. 31:45task requiring extensive source review,
  997. 31:47synthesis, architecture comparison, and
  998. 31:49likely iterative documentation or
  999. 31:51prototyping. So now, for example, I can
  1000. 31:53include a limitation here of please
  1001. 31:54limit your search to five sources and
  1002. 31:55the output document to 500 words or
  1003. 31:57less. And then I can reestimate it.
  1004. 31:59There we go. It just dropped from 45,000
  1005. 32:01tokens down to 18,000. There's a rough
  1006. 32:03estimate researching up to five sources,
  1007. 32:05comparing this particular task, and then
  1008. 32:06a synthesis under 500 words. And that
  1009. 32:08dropped the complexity of this task down
  1010. 32:11to medium. That's a cool way for you to
  1011. 32:12do a lightweight check ahead of time to
  1012. 32:14make sure that your task isn't going to
  1013. 32:15blow through your whole quota. So now
  1014. 32:16let's click create task. And this brings
  1015. 32:18up the board that I was talking about.
  1016. 32:21The triage section is where raw ideas.
  1017. 32:24We have a specifier that's going to
  1018. 32:25flesh out the spec. So it's going to be
  1019. 32:27converted from this into a different
  1020. 32:29card. And this is using that auxiliary
  1021. 32:31model that I was talking about. Okay,
  1022. 32:32here we go. So it just started. So that
  1023. 32:34triage using the auto decomposer, the
  1024. 32:36dispatcher checked. Oh, it's been a
  1025. 32:38minute. There's something here. Let's
  1026. 32:39break it into two different tasks. So
  1027. 32:41the task decomposer broke it up and
  1028. 32:43assigned it to these particular sub
  1029. 32:44aents where each one is now running on
  1030. 32:47its own. And we can click into this and
  1031. 32:48we can see the specific prompt that was
  1032. 32:50given to it created by the auto
  1033. 32:52decomposer using the default model. We
  1034. 32:54can estimate the effort of this
  1035. 32:55particular task like I was just showing
  1036. 32:56you. And we can comment. So as these
  1037. 32:58tasks are currently running, we can
  1038. 33:00actually send messages to the worker and
  1039. 33:02if there's a big enough change, we can
  1040. 33:03even send a message and then reue it and
  1041. 33:05have it start that task over again. So
  1042. 33:07we get a chance to go through and see
  1043. 33:09how all of this is working, what it's
  1044. 33:10actually doing, how it's analyzing it.
  1045. 33:12And we can see this for each card that's
  1046. 33:14running here. So you get a lot of
  1047. 33:15flexibility for controlling these cards
  1048. 33:17here. And what's cool, too, is for
  1049. 33:19example, I can go back to the to-do
  1050. 33:21section, which are the currently blocked
  1051. 33:22tasks. So these ones are waiting for the
  1052. 33:25researcher to finish before they run.
  1053. 33:27And we can see this one was actually
  1054. 33:28assigned to default instead of
  1055. 33:30orchestrator. So I'm going to change the
  1056. 33:31assigne from default here back to
  1057. 33:33orchestrator. And now we can see that
  1058. 33:34changed the card over here as well. So,
  1059. 33:36I think it's just such a cool way for us
  1060. 33:38to be able to take a look at how all of
  1061. 33:40the tasks are broken up into their
  1062. 33:42individual actions and see how they move
  1063. 33:44through the conbon and you can kind of
  1064. 33:46just get a sense on how all of the
  1065. 33:48project is working together. And that's
  1066. 33:50just for one card. I can now add a
  1067. 33:52second card and it can start running
  1068. 33:53itself and keep going from here. So, you
  1069. 33:55can really start to plan out maybe you
  1070. 33:57have five or six or 10 or 20 tasks that
  1071. 33:59you want to run. You can put them all in
  1072. 34:01here. You can schedule them. You can let
  1073. 34:02the agent know that some are waiting on
  1074. 34:04others. There's a lot that you can do
  1075. 34:06here and it just takes a lot of that sub
  1076. 34:08agent delegation that we were talking
  1077. 34:10about in the first couple examples and
  1078. 34:12makes it a lot more transparent and
  1079. 34:14controllable and that's why it's called
  1080. 34:15a control graph or a task graph because
  1081. 34:17we can actually see the graph of how
  1082. 34:19this is operating. What needs to be done
  1083. 34:21first before the next task can be
  1084. 34:23accomplished. So this one is still
  1085. 34:25running here but we can see this one
  1086. 34:26finished. We can go through and see the
  1087. 34:28full activity. So we can see that it was
  1088. 34:30created and then promoted to ready to
  1089. 34:31run and then it was claimed by a worker
  1090. 34:33which is this researcher agent. It ran
  1091. 34:35through completed it and it created a
  1092. 34:37briefing doc that was saved right here.
  1093. 34:39It only reviewed two sources because it
  1094. 34:41was given a two source limit and it
  1095. 34:44maintained everything. Okay. And it
  1096. 34:45looks like the second one just finished
  1097. 34:46as well. And now we can see that the
  1098. 34:48orchestrator task just moved from to-do
  1099. 34:51to ready. So once it's going from ready,
  1100. 34:53once the dispatcher ticks in 60 seconds
  1101. 34:56or so, it should get put over to
  1102. 34:58running. And there we go. So basically
  1103. 35:00we had those two research tasks and once
  1104. 35:02they were finished the to-do of
  1105. 35:03delivering the report was unblocked
  1106. 35:06because the previous tasks were
  1107. 35:07completed and then once that was
  1108. 35:09unblocked the dispatcher running in the
  1109. 35:11background moved it over to ready. And
  1110. 35:13then once it was ready it shifted it
  1111. 35:15into running. And again we can see all
  1112. 35:16of this was done by the auto decomposer.
  1113. 35:18And here's the specific prompt. So it's
  1114. 35:20pretty complex here. This is where it's
  1115. 35:22important to be really strategic with
  1116. 35:24the information and the tools and the
  1117. 35:26memory and the context, the skills and
  1118. 35:28everything that you're giving to your
  1119. 35:29specialized agents because you don't
  1120. 35:30want to accidentally have them blow
  1121. 35:33through your entire quota because you
  1122. 35:34gave them more than they needed. So
  1123. 35:35that's one of the dangers of running all
  1124. 35:37of this autonomously is that you might
  1125. 35:38be giving these agents more than they
  1126. 35:40need. So it's up to you to go through
  1127. 35:42and experiment and test and just make
  1128. 35:44sure that you're able to be more
  1129. 35:46specific and strategic with how you want
  1130. 35:48to set up these task graphs and how you
  1131. 35:50want to run them. So that task just
  1132. 35:51completed right here and now we have
  1133. 35:53multi-agent research ready. So we had
  1134. 35:54the researcher run create its research
  1135. 35:57report pass it off to the orchestrator
  1136. 35:59who then constructed a 500word or less
  1137. 36:02recommendation and then pass it back to
  1138. 36:04the original prompt that I had put in
  1139. 36:06here called multi- aent research. And
  1140. 36:08now this should be answered by it
  1141. 36:10passing along the document that was
  1142. 36:12created. Great. There we go. So now the
  1143. 36:13dispatcher is running again. The multi-
  1144. 36:15aent research my original question that
  1145. 36:17was created by me. And now it's going to
  1146. 36:19give me the report.
  1147. 36:21Great. And everything is now done. And
  1148. 36:23this is the report that it created from
  1149. 36:25incoming information to action and
  1150. 36:27reusable knowledge which is the theme
  1151. 36:28that we've been working with today. So
  1152. 36:30you can see it's 484 words. It is
  1153. 36:32limited to five sources and it gave a
  1154. 36:34pretty solid recommendation. Have a
  1155. 36:36single accountable lead agent with
  1156. 36:37bounded delegation is the best default.
  1157. 36:39Use a manager style orchestration to
  1158. 36:41keep one owner for state and a final
  1159. 36:43output while allowing scoped
  1160. 36:45specialists. Use a persistent team only
  1161. 36:47when measurement shows sustained
  1162. 36:48independent parallel work. So again,
  1163. 36:50this is basically the same answer as
  1164. 36:52what we had before. So now let's quickly
  1165. 36:54break down what just happened here
  1166. 36:55before I explain the alternative option.
  1167. 36:57So this is the original prompt that I
  1168. 36:59gave and what happened is you can see
  1169. 37:01this was created by dashboard. So this
  1170. 37:03was created by me. I assigned it to the
  1171. 37:05orchestrator and I gave it this
  1172. 37:07particular prompt. Then we can see that
  1173. 37:08the auto decomposer commented and said
  1174. 37:10hey I decomposed I broke up this task
  1175. 37:13into these three subtasks. So these were
  1176. 37:15the two research tasks and then the
  1177. 37:17writing of the article task. The root
  1178. 37:19this question here this task will wake
  1179. 37:22once all the children complete. So it
  1180. 37:23took that description it broke it into
  1181. 37:25those three subtasks. It went through
  1182. 37:27and ran the researcher went into running
  1183. 37:30completed its two tasks here. sent it
  1184. 37:32back to the orchestrator to write the
  1185. 37:33article, which honestly in the future I
  1186. 37:35would probably have an implement agent
  1187. 37:36or a writing agent who's specialized
  1188. 37:38with this, leaving the orchestrator to
  1189. 37:40just deal with orchestration. So that's
  1190. 37:42one change I'm going to make. And then
  1191. 37:44it gave the summary here. So it accepted
  1192. 37:45the deliverable. The final is 485 words.
  1193. 37:48All the sources check out. And here's
  1194. 37:50the simple recommendation. So honestly,
  1195. 37:52that worked pretty well. But maybe we
  1196. 37:53didn't need two researcher agents.
  1197. 37:55That's a little bit of overkill for a
  1198. 37:57five source limitation. that probably
  1199. 37:59used up more tokens than what we needed
  1200. 38:01here. So this is where if we close this
  1201. 38:02and we go back to our orchestration
  1202. 38:04settings, we can start moving beyond
  1203. 38:06this auto decomposer.
  1204. 38:12So this is where instead we can turn it
  1205. 38:14off and we could let the orchestrator be
  1206. 38:16the one to decompose the triage tasks.
  1207. 38:18So why would you want to do that? Well,
  1208. 38:20my orchestrator has a specific soul
  1209. 38:22associated with it. If I go here and
  1210. 38:23click edit soul, I have very specific
  1211. 38:26instructions on how I want the
  1212. 38:27orchestrator to work, how I want it to
  1213. 38:29coordinate my team, who I want it to
  1214. 38:30work with. I can put limits, I can
  1215. 38:32connect specific memories. There's more
  1216. 38:34context that can be given here from the
  1217. 38:37orchestrator profile than just the auto
  1218. 38:39decomposer, which doesn't necessarily
  1219. 38:41have that same level of context. This is
  1220. 38:42just automatic mode. This is more of a
  1221. 38:44curated mode, and it doesn't even have
  1222. 38:46to go through triage. Instead, it can
  1223. 38:48just go straight into ready. But if we
  1224. 38:49want to use the orchestrator to operate
  1225. 38:51as the task decomposer, we need to give
  1226. 38:54it specific conbon skills. So if I go
  1227. 38:56over to orchestrator and then we go to
  1228. 38:58capabilities, we can see here if I go to
  1229. 39:00tools again and then to task, I've
  1230. 39:02turned off task delegation. So basically
  1231. 39:04what that means is I've removed the
  1232. 39:06ability of the orchestrator agent to be
  1233. 39:08able to create those delegated sub
  1234. 39:09aents. Instead, I want the orchestrator
  1235. 39:12to run the conbon board. So I can turn
  1236. 39:14on these tools here. And what that does
  1237. 39:15is if we go back over to conbon by
  1238. 39:17having auto decompose turned off and the
  1239. 39:19orchestrator turned on as the
  1240. 39:21orchestration profile with the conbon
  1241. 39:23tools the orchestrator will now break
  1242. 39:25apart the tasks and then assign it to
  1243. 39:27the particular profiles defaulting to
  1244. 39:29the default assign here. So automatic
  1245. 39:31decomposition buys you convenience. It's
  1246. 39:33really easy. You just type in the task
  1247. 39:35and click go and Hermes will break it
  1248. 39:36down for you. But creating my own
  1249. 39:38planner as the orchestration profile the
  1250. 39:40orchestrator gives me a lot more
  1251. 39:42deliberate control over the task design.
  1252. 39:44and the roles in the review. It's not
  1253. 39:46that one necessarily has to be lower
  1254. 39:48cost than the other. It depends on the
  1255. 39:49task, the complexity, and how much
  1256. 39:51control you want to maintain and how the
  1257. 39:53system operates. The real benefit is
  1258. 39:56continuity. I can return tomorrow. I can
  1259. 39:58see what's blocked. I can inspect the
  1260. 39:59handoff, and I can resume without
  1261. 40:01reconstructing the conversation or
  1262. 40:03trying to set off more sub agents. So,
  1263. 40:05the orchestrator plans, the dispatcher
  1264. 40:07starts the system saying, "We're ready
  1265. 40:09for work." And then the worker executes
  1266. 40:10on the tasks assigned to it. So you can
  1267. 40:12kind of think of each card as work
  1268. 40:14units, not necessarily different agents,
  1269. 40:16but a card can use helpers or contain a
  1270. 40:18goal loop in and of itself. So if we
  1271. 40:20were to do that again, we set which
  1272. 40:22orchestration profile we want, what
  1273. 40:23default assign we want, but instead of
  1274. 40:26creating a new card in triage here, new
  1275. 40:28task in triage, it would just sit here
  1276. 40:29forever because the dispatcher doesn't
  1277. 40:32have the auto decomposer to break that
  1278. 40:35single task into the series of subtasks
  1279. 40:37that then would be put into ready. I'm
  1280. 40:39not going to go through that
  1281. 40:40orchestration example. The key
  1282. 40:42difference is that rather than dropping
  1283. 40:43something into triage, you can either
  1284. 40:45put it into to-do and then wait until
  1285. 40:46you're ready. You can schedule it or if
  1286. 40:48you want to trigger the orchestrator
  1287. 40:49right away, you would drop the new task
  1288. 40:51in ready and let it deal with the task
  1289. 40:53decomposition instead of letting the
  1290. 40:55auto decomposer do it. But I actually
  1291. 40:56tested out the different scenarios to
  1292. 40:58see which works best. So why don't we
  1293. 41:00take a quick look at the results of my
  1294. 41:02testing?
  1295. 41:07So my thought was there's a lot of
  1296. 41:08different coordination methods here.
  1297. 41:10which one actually makes the most sense.
  1298. 41:11The results showed that a single agent
  1299. 41:13with task delegation for a simple task
  1300. 41:16at least and running the orchestrator
  1301. 41:18without the auto decomposer actually had
  1302. 41:20pretty similar token usage and pretty
  1303. 41:22similar quality. But using the auto
  1304. 41:24decomposer and adding the goal loop
  1305. 41:26actually added more overhead for this
  1306. 41:28particular task. So this is obviously a
  1307. 41:30very simple example. It's not by any
  1308. 41:32means a proper benchmark. I just wanted
  1309. 41:34to see how this worked for a particular
  1310. 41:36research task for me. And one thing to
  1311. 41:38keep in mind is that even if the
  1312. 41:39workflow in the conbon takes more
  1313. 41:41tokens, it might be worth it because you
  1314. 41:43can go through and take a look at what
  1315. 41:45was actually done. So it gives you more
  1316. 41:46transparency. The actual right choice
  1317. 41:48just depends on what your work really
  1318. 41:50needs. So I'm going to keep exploring
  1319. 41:52this more. I'm going to try and do some
  1320. 41:54more complex conbon systems with coding
  1321. 41:56so I can actually see what makes sense
  1322. 41:58for my own workflows, how we can
  1323. 42:00leverage the orchestration and control
  1324. 42:02graphs that Hermes has built into it.
  1325. 42:06So, do you need an agent team? Not
  1326. 42:09necessarily. Start with the simplest
  1327. 42:11level that actually meets your needs and
  1328. 42:13then add coordination and complexity as
  1329. 42:15you find yourself actually needing more.
  1330. 42:18You want to find a specific problem that
  1331. 42:19that extra coordination and effort is
  1332. 42:22going to actually solve. Now, today's
  1333. 42:24example was a very simple one. Next,
  1334. 42:26we'll get more into how sub aents,
  1335. 42:28loops, and graphs can fit together to
  1336. 42:30solve more complex workflows. What
  1337. 42:31workflow would be the most helpful for
  1338. 42:33you? Please let me know in the comments
  1339. 42:34and I'm happy to explore these topics
  1340. 42:36more in future videos. If you found this
  1341. 42:38video helpful, please like and subscribe
  1342. 42:39as I really appreciate it. And please
  1343. 42:41consider joining my memberships. My
  1344. 42:42members enable me to continue making
  1345. 42:44free videos like this one. Thanks again
  1346. 42:46for watching and I will see you in the
  1347. 42:47next video.

About this transcript

This page contains the full transcript of Subagents vs Agent Teams? 🧠 Hermes Bots, Goal Loops & Kanban Graphs by Wanderloots, generated from the public captions YouTube serves with the video. The transcript has 9,470 words across 1,347 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.