YT YouTube2Text

YouTube2Text → Transcripts → Intelligence EXPLOSION: Harness Engineering with Pi Agent, Deepseek, and Gemini

Intelligence EXPLOSION: Harness Engineering with Pi Agent, Deepseek, and Gemini — Transcript

by IndyDevDan · 5,707 words · 848 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00What's up engineers? Andy Deb Dan here.
0:02Kimmy K3 was released mid-July, Deep
0:05Seek V4 flash late July early August,
0:07Quinn 3.8 early August, Muse Glimmer,
0:10Nemo Tron 3.5, Grok 4.6, Deep Seek V4
0:13Pro, Gemini 3.7 flash, Quinn 3.8 27
0:16billion, and GLM 5.3. We have just
0:20experienced an intelligence explosion.
0:22We have more than five model releases
0:25within five days across all tiers of
0:28models. The big AI labs are clearly
0:31panicked. Fable is a permanent part of
0:34the Claude subscriptions now, and OpenAI
0:36recently slashed prices for Tara, Luna,
0:39and now they're testing 50% cuts on GPT
0:425.6 sold on open router. The LLM pricing
0:45wars are in full effect. The
0:48intelligence explosion emphasizes two
0:51key [music] questions for us engineers.
0:53What's the best way to use these models
0:55together to outperform them
0:56individually, and how can we build for
0:59an age of rapid change so we can
1:02leverage the models with the best
1:03performance, speed, [music] cost
1:05tradeoffs for our valuable engineering
1:08work. I've been engineering for over 15
1:09years, and in technological revolutions
1:12where rapid change is the norm, one
1:14[music] engineering principle stands far
1:17above the rest. The most flexible system
1:20wins. [music] Today, I want to share my
1:22thoughts on the intelligence explosion
1:24and share a V2 of one of my favorite
1:26custom [music] pie coding agents, the
1:28fusion harness. We're going to talk
1:30about why the fusion harness is so
1:32valuable, and then upscale all of it to
1:34talk about in-loop agent decoding
1:36[music] and out-loop agent decoding. If
1:39you want to leverage these ideas for
1:40your engineering work, stick around
1:42[music] and let's capitalize on this
1:44massive intelligence explosion.
1:52There are three commands I want to share
1:53with you to give you concrete ideas on
1:55how to use flexible agent harness more
1:58effectively. Now, I recommend the pie
2:00coding agent, but feel free to use or
2:02build any tool that gives you the
2:04capabilities you're going to see next.
2:06Before I showcase these, we need to
2:08address the intelligence explosion
2:09because there's a lot here and there's
2:11never been more opportunity for us
2:13engineers thanks to this wide selection
2:15of compute. I maintain a running model
2:18stack of the best models. Here's all the
2:20new models that were just released over
2:22the past month and then of course I have
2:23my favorites. give you my thoughts on
2:26the current model landscape. Starting at
2:28the floor here with the lightweight
2:29models that can run on your machine.
2:32Music Glimmer was released. Meta is
2:34still in the race. We'll see how they
2:36perform in the upcoming model releases.
2:38Neutrino 3.5 lightning really really
2:41great release coming out of Nvidia. I
2:42want to see where this goes. Then we
2:45have the incredible Quint 3.8 27
2:47billion. I'm really really happy about
2:49this model release, but I'm also looking
2:50for that mixture of experts release. I'm
2:53curious if they're going to release
2:54something like this or if we'll have to
2:55wait for another version. Then we pass
2:57the B tier and move right into our daily
3:00workhorses. This is where a lot of the
3:02leverage is being created right now for
3:05engineers deploying language models into
3:07their engineering agents and their
3:09product agents at scale. We have the new
3:12GLM 5.3. I am picky and stingy with my
3:15model providers, so I'm waiting for this
3:17to come to a US based backed company. We
3:20of course had the incredible DeepSeek V4
3:22flash release. This model is absolutely
3:25cracked. The pricing is cracked even
3:26after they increase it, it's still going
3:28to be an incredible model to work with
3:30and it's in A tier with the DeepSeek V4
3:32Pro model that was just released. We're
3:33going to look at this one today and
3:35compare it side by side the incredible
3:37Gemini 3.7 flash. Out of all the models
3:41that were released in the intelligence
3:42explosion we're experiencing right now,
3:44this is my favorite model. Again, you'll
3:46see why in this video. Then we had Quint
3:483.8 and Grok 4.6 in addition to Kimiko
3:523, these are two models entering the S
3:55tier that are open weights. Okay, so
3:57obviously Grok not open weights, but
3:59Kimiko 3 and Quinn incredible models. Of
4:01course, it's basically impossible to run
4:03one of these unless you have tons of
4:05capital, but through model providers
4:06like my favorite fireworks through open
4:09router, these models are now available
4:11and they're fractions of the price of
4:13the current state of the art, the Fable
4:155, the Opus 5, the GPT 5.6 Soul. So, a
4:18lot of great intelligence to work with.
4:20Let's jump to the fusion harness and let
4:22me really highlight two of my favorite
4:24models in the intelligence explosion and
4:26we'll work through them to see how they
4:27compare directly with Fable 5. Open up
4:29the terminal. I've got Fable 5 running
4:31right next to GPT 3.7 Flash right next
4:34to Deep Seek V4 Pro. Let's see how this
4:36agent [music] team performs side by side
4:39on three specific tasks.
4:46So, DuckDB, one of my favorite in memory
4:49in process databases, has just released
4:52a preview of their V2. When this
4:54happens, I sit down with agents to
4:56understand it, to grok it, and to really
4:58see what became available. What's new?
5:00This is a perfect use case for a fusion
5:02harness. We have multiple sources of
5:04compute, multiple sources of
5:05intelligence we can tap into to
5:08understand new tools, to understand new
5:09technologies, to help us harness them in
5:11the best way. So, let's take the V2 of
5:14the fusion harness we've discussed on
5:15the channel before and push it to the
5:16next level {slash} FH opinion. So, I
5:19want the opinion from all of my models,
5:21from all of my compute on one issue.
5:23Here's the problem we're going to run
5:24here. I'm going to fire this off. We
5:26want to understand what the first most
5:28important release coming out of DuckDB
5:30is. What's the most important thing we
5:32should focus on for a developer building
5:34a local first analytics application
5:36should test first. So, you can see here
5:38our agents are getting to work. Gemini
5:39Flash is already completed. Deep Seek V4
5:41Pro getting to work here, 33 tokens per
5:43second so far. And then of course, the
5:45behemoth, the monster, the best model
5:47available, Claude Fable 5 completed its
5:50work there. Already you can see the
5:52massive difference in token costs from
5:54just one simple prompt. And here you can
5:56see our models completed. So, a couple
5:58things to note right away. When you're
5:59using models, there are three things you
6:02should be paying the most attention to.
6:03Performance against whatever you're
6:05trying to get them to do, speed, and
6:07then cost. We have that all right here.
6:10Performance, speed, and cost, context
6:12windows, model aliases so that they
6:14don't start competing against each other
6:16by understanding the model name. This is
6:18a very, very subtle thing you'll notice
6:20when you're using these models together
6:21a lot. You can never reveal the name of
6:24the model to the other model, otherwise
6:25they'll start emitting weird behavior.
6:27We'll talk about that more in a moment.
6:28Gemini 3.7 Flash is moving like the
6:31Flash. Insane the quick model,
6:32DeepSeek-V4-Pro 80 TPS, Claude Fable 5,
6:35the slowest, but also the most powerful,
6:36but also vastly more expensive, in order
6:39of magnitude more expensive. In fact,
6:41this trend's going to continue
6:42throughout, but you can see here the
6:43Fusion Harness is presenting us with the
6:45results of all of our models. What does
6:48this mean? Why is this valuable? We have
6:50three unique perspectives on the
6:52results. So, what is the most important
6:54feature we should pay attention to from
6:56the release of this new tool based on
6:58our suite of intelligence. The most
6:59important thing we should focus on is
7:01the new variant. We can now use variant
7:04to gain automatic structural
7:05decomposition at right time from our
7:07in-memory doc databases. This is the
7:09choice from Fable 5. This is the choice
7:11from Gemini Flash. And you can see here
7:13DeepSeek-V4-Pro actually drifting a
7:16little bit. Interesting. Wonder why that
7:18is. They all wrote out a little quick
7:20experiment on how to actually execute on
7:23the task that they recommend. So, this
7:24is a very, very simple prompt. And
7:26you've seen this before in many
7:27different ways and configurations.
7:29Instead of just running one model to
7:31give you feedback on something, run N
7:33models. And then push it a little bit
7:35further. Understand what it costs to
7:36deliver that single unit of result,
7:39because guess what? That's going to
7:40scale as you scale your efforts with
7:43your agents. [music]
7:49Let's move to a much more powerful
7:50prompt that gets out of what everyone
7:52else is doing {slash} Fusion Harness
7:55debate. Inside of the DuckDB
7:57announcement post, they said something
7:59really interesting here. You can now use
8:00DuckDB as a server with Quack and
8:03connect. They broke down exactly how you
8:05can do this very cleanly and concisely.
8:07This is really interesting because now
8:09it questions how we should best use
8:11DuckDB. Is this a simple in-memory
8:14analytics processing tool that optimizes
8:16on the columns or is this more like a
8:18server that we should connect multiple
8:19instances to? We don't know. Let's ask
8:22our team of intelligence. What do they
8:24think about this? That's exactly what
8:25this prompt was. Debate this claim using
8:28the DuckDB V2 preview announcement. And
8:30then here's the claim as a single quote.
8:32And now they're going to start debating.
8:34So this is a different agentic workflow
8:36built into the fabric of my pie coding
8:39agent as an extension. This is behavior
8:41you're not going to find in any
8:43out-of-the-box agentic coding tool.
8:44We're having multiple models debate in
8:46several rounds. So let's walk through
8:48this. And it's not just one model, it's
8:50three models. And in fact, this system
8:51goes up to five models, but you can see
8:53round one just completed. They all gave
8:55their concrete opinion on the question
8:59we presented. Debate this claim. You can
9:01see that Fable says the claim is wrong.
9:03You can see from Gemini 3.7 Flash, team
9:05should continue treating DuckDB as a
9:07primary embedded analytical engine. And
9:09DeepSeek V4 Pro's position is that we
9:12should use this as a pilot, as a test.
9:14We should not move this to a production
9:16default general-purpose multi-tenant
9:18server. Every single model is converging
9:20on the same answer here. This is
9:22important. This is valuable information.
9:24Now, they all spent different amounts of
9:26time and tokens on this as we can see
9:28here. Of course, our big offender being
9:30Fable. That one response cost me 15
9:33cents. Definitely drop a like.
9:35Definitely drop a subscribe so uh
9:37YouTube can pay for this. But, if we
9:38scroll down here, you can see that the
9:40next round has started. So, my agents
9:42aren't just giving me a single opinion.
9:44They're not just saying something,
9:45they're debating. Okay, so let's scroll
9:46down, and let me show you exactly what
9:48this looks like, right? This is round
9:49two, updated side. So, what happened
9:51there? Every agent shared its response
9:53with every other agent in the system.
9:55And so, we get this type of structure.
9:57If we look top down, right? Everyone
9:58shared their opinion, and now they're
10:00going to the next round of debate. You
10:02can see how this can be very valuable.
10:05unique perspectives and diverse opinions
10:07coming out of our language models
10:09compete, compare, contrast around a
10:12central idea, a central topic of debate.
10:15So, you can see here, no one is budging.
10:17Unchanged. It makes it a credible
10:18network analytic endpoint worth
10:20piloting. Okay, Opinion Map, there is no
10:22pro side. Position here, unchanged.
10:24Position unchanged from Gemini 1.5 Flash
10:27and from DeepSeek V4 Pro 0813. And then,
10:30here's the cool part, you can see the
10:32agents taking in the opinions of the
10:34others. So, we have our Rune agent, Flux
10:36agent, Drift agent hiding the model name
10:39so that these agents don't get weird and
10:41start competing and start sabotaging
10:43each other. Something that just emerges
10:44naturally. I'm not sure what that's
10:46about, if that's part of the training
10:47data, or how that emerges. But, it's
10:50something important to note. If we keep
10:52scrolling down here, you know, you can
10:53see refutations, agreements, what
10:55changed my mind? Nothing did. You can
10:57see that they're basically just
10:59reinforcing each others' decision now.
11:01And then, we have our final closing
11:03statements. So, the debate is ending.
11:05Everyone gives a final answer. Reject
11:06the claim, reject the claim, reject the
11:08claim. And of course, we've prompt
11:10engineered the system so that they have
11:12a nice structured debate, and it all
11:13makes logical sense. And that ends our
11:15debate. There's the total cost for the
11:17entire session. This is what we're
11:19spending so far. Quick call outs here,
11:21when you're comparing these tools side
11:23by side, you can very, very quickly see
11:25how fast are these models, and who is
11:27performing the best. So far, performance
11:29is dead even, and you can see, of
11:31course, the costs. Part of what's going
11:33to happen here, as you already know,
11:35we're just going to keep emphasizing how
11:36expensive these state-of-the-art models
11:38are versus our A-tier compute. The
11:40Intelligence Explosion is really giving
11:43us great A-tier workhorse compute that
11:45is competitive for most tasks with the
11:48state of the art. That's a really,
11:50really big driver that's happening right
11:52now. We're getting costs down. I care
11:53less about cost down and I care more
11:56about effective agent hour per token
11:59cost. That's what really matters here.
12:01And in this A-tier, we're getting a lot
12:02of heavy hitters, specifically in this
12:04top row. The benchmarks are really,
12:07really spelling success for this top
12:09row. The gap here is really large on a
12:11cost basis. If you look at Gemini 3.7
12:13Flash, less than a dollar and four bucks
12:16out. Deep Seek before Pro going even
12:18further beyond that. And then we look at
12:20these state-of-the-art model prices, it
12:21gets really, really hard to keep using
12:23these unless you're inside some
12:25subscription plan that is kind of
12:27keeping you in a nice bubble defended
12:29from the true costs of these models. A
12:31lot of people don't know that there's
12:33some hidden text within a lot of these
12:35models. For instance, GPT 5.6 series,
12:38once you pass that 280K mark, the price
12:40doubles for your input and 1.5 access
12:43for your output. So, this model is not
12:45as cheap as it looks. If you compare
12:47Opus 5 and GPT 5.6 Soul, Opus 5 is
12:50better and it's cheaper. Now, just like
12:52in last week's video, there are some
12:53problems with Opus. We employed some
12:55serious system prompt engineering to fix
12:58a lot of the ticks to address how
12:59verbose, how many output tokens, and how
13:01load-bearing Opus 5 is. Check out that
13:04video, I'll link that in the description
13:05for you. If you want to master system
13:07prompt engineering and get more leverage
13:08and stop hyperfixating on your skills.
13:10But, you can see here that on many
13:12dimensions, not all models are created
13:14equally and that's what the model stack
13:15is about. We're combining compute, we're
13:17not selecting it. That's great. So,
13:19that's debate. We're putting models
13:21against each other against a specific
13:24thesis that we you to flush out more.
13:26This is great not just for learning new
13:27tools or technology or giving you
13:29direction. This is really, really great
13:31for decision-making. Strategic
13:33decision-making is one of the highest
13:35leverage things an engineer or PM,
13:37pretty much anyone can do. If you take
13:39an action that gives you 10x, 100x
13:41leverage, that can change everything.
13:43Some decisions you and I commit to,
13:45they're month-long decisions. They're
13:47year-long decisions. So, if you're going
13:48to put that much time in, you better be
13:50sitting down with some of the best
13:51intelligence available to you to help
13:53you make the decision or at least flush
13:55out the decision you're planning on
13:56making or giving you counterpoints. So,
13:58that's the debate command. This is
14:00easily my favorite command of the three,
14:02but the next one I'm going to share with
14:04you is the [music] most powerful.
14:11/fh collaborate. Now, our agents have
14:14context. We've been having these
14:16discussions, they're debating. Now, I'm
14:17going to say this, build a working demo
14:19of the top three features from the
14:21DuckDB release executable via Astro UV
14:24single file scripts. Fired it off, and
14:26now they're going to start a
14:27collaboration agentic workflow. Again,
14:30this is built into the tool. When I boot
14:32this agent up, I have a unique workflow
14:34that basically no one else has. It gives
14:37me outsize leverage on my agents. And
14:40therefore, more ROI on my engineering
14:42work. Now, of course, you're here,
14:43you're with me. Big shoutout to you. I'm
14:45going to link this tool in the
14:46description for you. You're going to
14:47have the V2 version with all these tools
14:49available to you. Definitely drop a
14:51like, subscribe, all that good stuff if
14:53this is valuable information and if this
14:55is helping you understand the landscape
14:57even just a little bit better. We're
14:58starting out with the same workflow. Our
15:00agents are looking at what's going on
15:02here. They're understanding the current
15:03project structure. Now, something really
15:05interesting is going to happen. As soon
15:06as Deep Seek V4 Pro finishes up here,
15:08all of our agents are going to present a
15:10plan. Once again, we're scaling our
15:12compute to scale our impact. We're not
15:13just getting one plan from one model.
15:15We're having every model in our system
15:18give us a concrete plan. And then, as As
15:20can imagine, we have our two builders
15:22here, and we have one architect. The
15:24Fusion Harness always has one architect.
15:26This is where you're going to want to
15:27put your most powerful model that you're
15:29willing to spend on, because the
15:31architect is what's going to take all of
15:33the opinions, all of the plans of our
15:35agents, and it's going to put them
15:36together into something incredible. So,
15:38let's look at the plans, right? Here's a
15:39proposed in state from every single
15:41model. We're getting multiple
15:42perspectives from our agents, Claude 3
15:44Opus 5, Gemini 1.5 Flash, Deep Seek v4
15:46Pro, and they're basically putting
15:47together a plan. Now, you can see here,
15:49very similar. This is a relatively
15:51simple thing to do. We want three astral
15:53UV scripts to showcase the three top
15:55features we should pay attention to from
15:57the DuckDB v2 preview release. They're
15:59breaking down their plan piece by piece,
16:01task dependencies, and then they're
16:02doing something interesting. They're all
16:04explicitly talking about their team, who
16:06should build what. And then we have a
16:07collisions and safety concerns. We've
16:09basically done a little bit of
16:10templating with a simple system prompt
16:12to help guide our agents create their
16:14own plan. So, if you watched last week's
16:16system prompt engineering video, you
16:17might notice something really awesome
16:19here. We're utilizing that inside of the
16:21system prompt for all of our agents. You
16:23can see here, we have our reference
16:24points T1, T2, T3, that we can quickly
16:27reference between any one of our agents
16:29or all of our agents if we want to. You
16:31can see here, T1, T2, T3, different
16:33tasks. There's a risk analysis. We have
16:35our Rs, so on and so forth. And we have
16:37clear language, we're clearly
16:38communicating with our agent. There's no
16:40load-bearing garbage in here. And again,
16:42I'll link last week's video in the
16:43description for you. If you want to
16:44understand how we fixed Smart Ass Opus 5
16:47by recentering on the most important
16:49skill any engineer can learn, prompt
16:51engineering. The simplest things often
16:53have the greatest value if you do them
16:55right. Prompt engineering is one of
16:57them. So, our architect agent is
16:59creating a simple task list, where you
17:01build and there's a dependency on
17:03previous tasks. You're very familiar
17:05with systems like this. The great part
17:06about this is that this is getting built
17:08on the fly, and our architect agent
17:10assigns an owner and a mode for each one
17:13of these tasks. And we have task
17:14dependencies, classic stuff. If we
17:16scroll down here, you can see our agents
17:18are just getting to work. And we've
17:19built in our till done lists that we've
17:21talked about in the past. It's a to-do
17:23list for our agents, except we are
17:25explicitly assigning tasks and the
17:27dependency workflow for each task. So,
17:30this lets our models collaborate on the
17:32solution and deliver a unique solution
17:34using many, many models. I'm using three
17:36models here, mostly for simplicity. We
17:38can see these three models here side by
17:40side by side, but there are versions you
17:42can run here. If we open up a new
17:43terminal here, you know, I have J Fusion
17:455 where we have five models put
17:47together.
17:48>> [laughter]
17:48>> And we could do the same thing with
17:50other models. You can scale this up or
17:52down as much as you like. The whole idea
17:54here is you want to combine compute. You
17:56don't want to select compute. You want
17:57to use the best model for the job, and
17:59you want to use a tool like the PyCoding
18:02agent that's extensible that lets you
18:03customize the very experience of your
18:06agentic engineering. And so, as I'm
18:08working through, testing, validating,
18:11building with models for in-loop agentic
18:13coding and out-loop agentic coding, I'm
18:15constantly comparing these models to
18:17truly understand what does a token here
18:20give me? How fast is that token come
18:22out? And how valuable is that token
18:23versus the cost of the token? Once
18:25again, same theme, nothing new. Fable,
18:2865 cents for that run. Gemini 3.7 Flash,
18:30just 7 cents. And DeepSeek V4 Pro, just
18:335 cents for that run. So, really
18:35important stuff. You can also see here
18:37something really fascinating. Gemini 3.7
18:38Flash is likely the fastest, most useful
18:41model in existence right now. It is just
18:44insanely fast. You know, a lot of people
18:45are saying Gemini, Google, they're
18:47falling off the AI race. They can't
18:48produce a pro model. They're not
18:50competing anymore. When I look at Gemini
18:513.7 Flash, I see a strategic decision to
18:54optimize for speed, cost, and
18:56intelligence and to say, "We don't need
18:58to be at the frontier to be relevant
19:00here." And I don't know about you, but
19:02when I look at this and I use this
19:03model, they're 100% right and they're
19:05doing it. I of course have private
19:06benchmarks for these models and I sit
19:08down like this and run these models
19:10together side by side by side to really
19:13understand what's happening. So, you can
19:14see tasks getting completed here one by
19:17one. Fables done, Flash is done,
19:19DeepSeek v4 Pro getting to work here.
19:21One thing I have noticed about DeepSeek
19:22v4 Pro, it thinks a lot. This is one of
19:24the heavy deep thinking models, very
19:27similar to Kimiko 3. Some of the other
19:29models like the new local model Qwen 3.8
19:3227B, great model, also thinks a lot and
19:35is thinking a lot without mixture of
19:37experts. And it really, really hurts the
19:38performance and the time to get those
19:41tokens back on your local machine.
19:43Obviously, throw this on a GPU node
19:44somewhere, cluster somewhere and it runs
19:46a lot faster, but when we're thinking
19:47about lightweight models, models we can
19:48own running locally, this is where the
19:51Qwen 3.8 model has been a little
19:53disappointing for me. It's performing
19:54really, really well. If we open up
19:56artificial analysis, it's an insane
19:57model that's getting above that 50 mark,
20:00but when I sit down to use this thing,
20:02the context window becomes a serious
20:04problem. So, all that being said, it's
20:05incredible model. I'm using this model
20:07locally, but we're absolutely going to
20:08need some performance speed ups on this
20:10model, which I am very, very confident
20:12they're coming. It's pretty crazy that
20:14at least on this benchmark, you can't
20:15take benchmarks as law, as truth for
20:17everything. Really, you need to deploy
20:19these models against your specific use
20:20case to really know if they're valuable,
20:22but just as input to your decision, we
20:24have an open weights model that you can
20:26put on your device that's running at the
20:28performance of GPT 5.6 Luna Max. So,
20:31very, very powerful. It's also
20:32apparently one intelligence point away
20:34from GLM 5.2 Max. This is where things
20:37kind of fall apart. Is it really that
20:38good? Doesn't really feel that good.
20:40Always take the intelligence benchmarks
20:43with a grain of salt and build great
20:45extensible flexible tools where you can
20:47compare them yourself. You can see here
20:48our agents are working through that
20:50checklist. What is this? This is
20:51multi-agent orchestration. What I've
20:53just showcased here are three different
20:55multi-agent orchestration patterns you
20:57can deploy into the fabric of your agent
20:59harness. Big fan of Cloud Code, still
21:01using Cloud Code. This subscription is
21:02one of the big pieces keeping me
21:04continuously using Cloud Code, but there
21:06are things that you just can't do when
21:08you're locked into someone else's agent
21:10harness. When you're locked into someone
21:11else's product, that's by design, but
21:13it's also going to hold you back,
21:15especially if you're an engineer that is
21:17pushing into the new role where we are
21:19still software engineering, but with
21:21autonomous technology that can take
21:23actions on your behalf. If you're really
21:25pushing into that, the agent harness is
21:27the thing to own. And if the tool you're
21:29using, closed source or open source,
21:30isn't customizable enough, isn't
21:32flexible enough, doesn't let you change
21:34as models continue to change, you're
21:36going to run into some roadblocks that
21:37you don't need to. And other engineers,
21:40like myself and like many other
21:41engineers that follow the channel, these
21:42are just not problems we're going to run
21:44into because we have both a closed
21:46source option and an open source super
21:48flexible option. I totally get cost
21:50optimizing. I totally get using out of
21:52the box models, especially for
21:53enterprises, like you really need a lot
21:55of extra stuff. But for experimenting,
21:57for really understanding the landscape,
21:59you really, really, really want to own
22:01your agent harness. And then the next
22:03step is to own your model. More on that
22:06later. Obviously, that's a much harder
22:07problem to solve. We talked about that a
22:09lot in the is Anthropic stealing your
22:11data while you pay for it video. I'll
22:13link that in the description for you as
22:15well. But you can see here, our agents
22:16are just working through the checklist.
22:18It looks like right now Flux, which is
22:19our Gemini 3.7 Flash or lightning flash
22:22fast model, is appending the demo
22:24receipt to our just file so that we can
22:27quickly run this demo of all these
22:28features. Last but not least, our
22:30architect is responsible for the final
22:32integration. It's doing the validation.
22:34If you're in the AI industry, this is
22:36stuff you've seen before, but we're
22:37combining compute to do the job of what
22:39we would normally deploy one model for.
22:41We can deploy many, understand the
22:43model, and get work done, and have our
22:45compute check itself. We're adding more
22:47compute, adding more scale so that we
22:49can have more impact. On the model
22:51level, I highly recommend you check out
22:52Gemini 3.7 Flash. It's incredibly cheap.
22:55Big fan of DeepSeek V4 Pro. This is
22:57going to be the open weights model. If
22:59you have the hardware, if you have the
23:00technology to host this and also
23:02fine-tune this. This is going to be a
23:04fantastic model, much like Deep Seek V4
23:06Flash. Even as I look at all of these
23:08models, quote-unquote, catching up, when
23:10I sit down to do hardcore advanced
23:13agentic engineering work, I'm still
23:15orchestrating with the Claude Fable 5
23:17model. I'm not orchestrating with Opus
23:205. I prefer Fable 5. Opus 5 is fine, it
23:23has a lot of problems. It's too hungry.
23:26It's looking for problems where they
23:27don't exist. And again, last week we
23:29covered how you can try to attack this
23:31in the system prompt. That works
23:32decently well, but you know, for a lot
23:34of the work I'm doing in loop, I am
23:36opening up Claude Code, throwing Fable
23:37at it, and then I'm using my many, many
23:39custom variants of my Pi coding agent to
23:41get the job done. And this is one of the
23:43variants that I'm putting more time
23:44into, because I'm getting a lot of nice
23:46results out of this model, out of
23:47multiple model perspectives. Again, as
23:50there is an intelligence explosion, it's
23:51going to become more and more important
23:53for you to understand each model. And
23:54this is a great way to understand
23:56multiple models at the same time. Here's
23:58our final result. Here's our final cost
23:59analysis. You can see here, it's about
24:01an order of magnitude cheaper to use 3.7
24:03Flash or V4 Pro over a state-of-the-art
24:06model like Fable. I know a lot of
24:07engineers watching the channel um that
24:09like to complain in the comment section,
24:10while you're really using the API, while
24:12you're really spending those tokens.
24:14Yes, I am. If you want outsized returns,
24:16you have to do things others aren't
24:17doing. This is one of them. One of the
24:19easy ways to spend capital to get an
24:21advantage. Of course, use your
24:22subscription. Of course, look for the
24:24good deal. Again, we talked about this.
24:25When you're using the API, you have
24:27higher levels of protection for your IP.
24:29There's real ZDR built in. So, anyway,
24:31but here is our Fusion Harness result.
24:33Completed. Our architect says
24:35collaboration is closed. Here is the
24:37result. You can see there's all the
24:38deliverables. If we open up VS Code, we
24:40can see exactly what was built in this
24:42code base. Here is our demos. You can
24:45see here, I also threw the system prompt
24:47from our Fixing Opus 5 video from last
24:49week right in here, gave it to all my
24:51models. We can look at the
24:51configuration. Super, super simple.
24:53Here's the one I was using for this.
24:55Append system prompt. You're going to
24:56append multiple system prompts, pretty
24:58nice. But then, just model name
25:00thinking, keep it simple, don't
25:01overcomplicate things. And here is our
25:04demos that our agents built showcasing
25:06the new DuckDB work. Good stuff here.
25:08I'm not going to run this. Obviously,
25:10this isn't the point of the video. This
25:11is more for my personal exploration of
25:13DuckDB. I'll go ahead and leave the
25:14demos in this codebase if you're
25:16interested. I'll port it over to the
25:17full Fusion Harness codebase, which is
25:20available to you, link in the
25:21description for the full breakdown, for
25:23the full Fusion Harness. This is the V2
25:25version. We did a video on this a few
25:27weeks ago. I really just want to stress
25:28that idea. You want to think in ands,
25:30not ors. It's not GPT-5.6 sole versus
25:33fable. You want to use them together.
25:35You want to stack up compute, combine
25:36compute, don't select compute. That is
25:38the catchphrase for this tool right at
25:40the top, combine your compute. There are
25:42many other commands I did not run here.
25:45I'll leave this as exploratory work you
25:47can do if you're interested in the
25:48Fusion Harness. Auto Validate is a big
25:50one. The Fusion command is another big
25:52one. So, link in the description for
25:54you. Think about the leverage that's
25:56stacking up versus an engineer using a
25:58tool like this versus using some
25:59out-of-the-box tool. They run three
26:01models, they get three concrete
26:02opinions, they get three different
26:04builds, they get three different plans.
26:05Then they combine it or they select the
26:07best. There's so many ways to build
26:10flexible pi coding agents, build
26:11flexible tools, and then use these
26:13models together to outperform the normal
26:15distribution. More and more and more
26:18engineers are using out-of-the-box
26:19tools. Once you understand that you can
26:21harness engineer your own specialized
26:24units surrounding your agents, then
26:27comes the next level that we've been
26:28talking about on the channel week over
26:30week, the software factory. This is
26:32where you combine agents plus code to
26:34outperform either alone. This is where
26:37the next generation leverage is. Very
26:39few engineers are really pushing toward
26:41this edge. This is where the real alpha
26:43is now. It's not about you sitting in a
26:45terminal, babysitting your agent,
26:46prompting back and forth over and over
26:48for every problem you have. This is
26:50about you using out-loop agentic coding.
26:52It's about stepping out the loop where
26:54you build software factories that work
26:56on your behalf without you. So, that's
26:58the next step of leverage. If you don't
27:00know where to focus on, if you feel like
27:01you're stalled, and you understand
27:02prompt engineering, context engineering,
27:05harness engineering, and if you
27:06understand that it's not loop
27:07engineering you're after, it's building
27:09and managing different variants of the
27:11software developer life cycle. If you
27:12get that, the next step is the software
27:14factory. You have to get out the loop.
27:16The technology is there. The models, as
27:19you can see very very clearly, are
27:21there. Now, it's about you taking that
27:23next step. Stop blaming the models. Stop
27:25blaming the tools. Everything is in your
27:27control. If you do this right, you can
27:29change everything. But, you have to put
27:30your foot forward. You have to take
27:31responsibility for [music] the results
27:33you're saying. The next level is the
27:34software factory. We're going to keep
27:36seeing more of these intelligence
27:37explosions. I'm going to keep tracking
27:39all these models. My goal here on the
27:40channel is to be a trusted source and
27:43input to your decision-making [music]
27:44process to help you push your agentic
27:47engineering far beyond the rest. Huge
27:49thanks from me to you [music] for
27:51watching to the end. Drop a like if you
27:53can see the opportunity the intelligence
27:55explosion is unlocking for all of us.
27:57You know where [music] to find me every
27:59single Monday. Stay focused and keep
28:02building.

About this transcript

This page contains the full transcript of Intelligence EXPLOSION: Harness Engineering with Pi Agent, Deepseek, and Gemini by IndyDevDan, generated from the public captions YouTube serves with the video. The transcript has 5,707 words across 848 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.