Merge Is the Bottleneck - Prasanna Sankar, Vorflux — Transcript
Full transcript
- 0:05Hey guys,
- 0:06my name is Prasanna and I really
- 0:08appreciate the opportunity to speak in
- 0:10front of all of you guys. Today,
- 0:12AI has really accelerated leverage and
- 0:15you know, this is a story of leverage.
- 0:17I'll start with my own personal
- 0:18background.
- 0:19I'm a software engineer,
- 0:21was ranked number one coder in India
- 0:24when I was 18 years old during my
- 0:26college days.
- 0:27I won [clears throat] all these
- 0:28programming contests like top coder. I
- 0:30was ranked number one there.
- 0:33I was two times world finalist in both
- 0:34Google Code Jam and ACM ICPC. These are
- 0:39the you know, the top programming
- 0:40contests in the world and did really
- 0:42well there, represented India
- 0:45quite a few times. Now
- 0:47[clears throat]
- 0:48the models have gotten way better than
- 0:50any of us. They perfect score these
- 0:52programming contests. So it's an
- 0:54interesting time to be alive and you
- 0:56know, super exciting time. I want to
- 0:58talk a little bit about you know, my own
- 1:00background, right? Like my own journey.
- 1:01After college, I kind of like worked at
- 1:03Microsoft for a year, quickly realized
- 1:05you know, the bottleneck to
- 1:08impact on the real world or my own
- 1:11personal leverage
- 1:13was other people
- 1:15like working in big teams.
- 1:17[clears throat] And then there was the
- 1:18the alternate path. I came across Paul
- 1:20Graham's blogs.
- 1:22Silicon Valley was happening.
- 1:24Internet companies were getting a lot
- 1:26cheaper to start. So a lot more people
- 1:28like myself
- 1:30were able to access it.
- 1:32And Paul Graham's blogs gave me a path
- 1:34towards living a life of higher
- 1:36leverage. So I moved to the valley
- 1:38sometime around 2009 to
- 1:41you know, start building companies. At
- 1:43this point, I'm a three times YC
- 1:45founder.
- 1:47[clears throat] The most successful
- 1:48company that I founded was a company
- 1:50called Rippling, which is worth more
- 1:51than 10 billion today. More recently, I
- 1:53started a company called Warflex. It's a
- 1:55it's a part of the latest YC batch as
- 1:57well. We've raised around 15 million
- 1:58dollars.
- 1:59And it's a AI software engineer
- 2:03or AI coworker, the highest leverage AI
- 2:06coworker. So the way I got that is like,
- 2:08you know, post founding Rippling,
- 2:11I was just like building things for
- 2:12myself and I was building things of like
- 2:15higher and higher leverage. Something
- 2:17interesting happened in in in my world
- 2:19with the launch of
- 2:21Opus 4.5. I think it was like almost a
- 2:23year ago at this point. It was a magical
- 2:26moment.
- 2:27Software went from autocomplete in in
- 2:29the IDE
- 2:31to
- 2:32[clears throat]
- 2:33agents working by themselves for a few
- 2:35minutes at a time. So you could suddenly
- 2:37sort of move from completely skip the
- 2:39IDE and move into the terminal. That was
- 2:41seminal, you know, it was naturally
- 2:43obvious that, you know, that was just
- 2:44going to take off. Those few minutes are
- 2:46going to get longer and longer and
- 2:48longer.
- 2:49So
- 2:50[clears throat] you know, at the same
- 2:51time, these models started winning all
- 2:53these programming contests and it was
- 2:54kind of like fairly clear to me like
- 2:56they were better than me in just like
- 2:59raw reasoning.
- 3:00Especially, you know, but then the the
- 3:03the harnesses that won these programming
- 3:06contests were not a vanilla Claude code
- 3:09or a Codex harness. They spent a lot
- 3:12more on inference time compute, a lot
- 3:15more inference time compute to like, you
- 3:16know, access higher order intelligence.
- 3:19So that's what I wanted to build for
- 3:20myself, you know, the higher leverage
- 3:22version which spends a bit more tokens,
- 3:24but you know, saves a lot of time, you
- 3:27know, gives me like a lot more
- 3:28intelligence.
- 3:30So I started building Worx Flex for
- 3:32myself
- 3:34and, you know, the first obvious thing
- 3:36is like, you know, once agents start
- 3:37running beyond a few minutes into into
- 3:41hours at a time, especially if you're
- 3:43able to solve a bunch of these
- 3:44bottlenecks that that I'm going to talk
- 3:46about,
- 3:47you really want them running on the
- 3:49cloud. You don't want them running on
- 3:50your machine because, you know,
- 3:53that's where they belong, a task that
- 3:56runs hours at a time.
- 3:57So [snorts] I think that's the first big
- 3:59bottleneck that people need to solve,
- 4:01like, you know, cloud agents need to
- 4:02solve to get
- 4:04people moving from the desktop to to the
- 4:06to the cloud is a perfect environment
- 4:09setup. Like today cloud agents run like
- 4:12most cloud agents run blind. At best
- 4:14they run the one service that you're
- 4:17trying to change, but really like for
- 4:19for like good testing the reason why
- 4:22people sort of like do all this stuff on
- 4:24their local host is because
- 4:26there are a lot of other dependencies
- 4:28and you want to be able to run you know,
- 4:30you want to be able to almost like
- 4:31mirror the local host environment on the
- 4:33cloud to truly transport someone. So we
- 4:36spent a lot of tokens and money and time
- 4:39solving the machine setup bottleneck
- 4:42where you can just like onboard and we
- 4:44start a bunch of sub agents to like go
- 4:46set up raw EC two instance
- 4:49with your whole environment, wire up the
- 4:51ports,
- 4:53figure out the machine secrets, solve
- 4:55authentication, solve seed data and a
- 4:57bunch of these things, spend tokens
- 4:59upfront migrating the user to the cloud.
- 5:02I think that's really important. That's
- 5:03something that most cloud agents miss.
- 5:05So that's that's like one of the major
- 5:07bottlenecks to moving from desktop to
- 5:08the cloud.
- 5:10Once you solve that, what you realize is
- 5:13like if you you know, when I was
- 5:15profiling how people were spending their
- 5:17time including myself, a lot of the time
- 5:18was being spent on the planning phase
- 5:21and
- 5:23you know, [clears throat]
- 5:24like
- 5:25back then the planning agents
- 5:29were not spending as much tokens.
- 5:32The agents weren't spending as much
- 5:33tokens on the planning phase. You know,
- 5:35they usually come back within a minute
- 5:37or two and you were iterating a lot back
- 5:39and forth, back and forth on the
- 5:40planning phase. And you know, obviously
- 5:42if human time is being spent there,
- 5:44that's where I want to throw more
- 5:45tokens, right? So what you know,
- 5:46workflows we sort of scaled up
- 5:50inference time computed the planning
- 5:51stage. We launched a bunch of parallel
- 5:53explorer sub agents which sort of built
- 5:57built a design or an architecture of
- 5:59what needs to be done
- 6:00and you know, we sort of came up with a
- 6:03plan of plans which are like a lot more
- 6:05detailed and you know and then sort of
- 6:09produced a summary out of it to help
- 6:12solve the time spent by humans on the
- 6:15planning phase, right? So that that was
- 6:18like a major bottleneck that we were
- 6:19trying to solve. Then comes
- 6:21orchestration. People usually often like
- 6:24break down like a large task into a
- 6:25bunch of small tasks and like work hard
- 6:28trying to land each one of them.
- 6:30We are like default sub agent build, you
- 6:33know, once a plan is built, you know, it
- 6:34has a dependent it comes with a
- 6:36dependency graph and sub agents go do
- 6:39independent pieces of work, right?
- 6:41And then testing is like another major
- 6:43bottleneck. If
- 6:47[clears throat] if you still need to
- 6:48pull stuff to localhost to test like
- 6:51like you you can't transition to the
- 6:53cloud. So we built
- 6:55like
- 6:56a detailed testing framework where you
- 6:59know
- 7:00browser testing happens on the machine,
- 7:03end-to-end testing happens and
- 7:06you know the output is something that
- 7:08the that proves to the user that things
- 7:10are working, right? So for me like I
- 7:13really enjoy like looking at a video
- 7:14which proves that the that the new
- 7:18feature actually works or you know,
- 7:20sometimes like I really you know, it
- 7:22also gives me like a staging link where
- 7:25for that particular feature I can click
- 7:27around
- 7:28because it's anyways hosted on a
- 7:30machine, right? So you can just like run
- 7:32the whole app on that machine, click
- 7:34around, play with it and
- 7:38and see if it works. So that's the
- 7:40that's the way to like completely remove
- 7:42localhost as a requirement for testing.
- 7:44Post that code review becomes a big
- 7:46bottleneck. Often times people ask
- 7:50multiple models to do reviews for you
- 7:52and we bake that in inside the harness.
- 7:55You can customize the harness and many
- 7:57people often ask often build with Claude
- 7:59and get GPT to do their review
- 8:04and you can have like multiple reviewers
- 8:06looking at like different
- 8:08These are all sub agents. Um, so you can
- 8:12basically customize the harness which
- 8:14um solves a lot of the reviewing
- 8:16bottlenecks.
- 8:17Then comes the merge. Um, often in fast
- 8:21moving teams when people are agents are
- 8:23opening a lot of PRs in parallel, um you
- 8:26need something like a merge queue which
- 8:27we built
- 8:29which constantly like rebases, keeps
- 8:32keeps you mergable with master and like
- 8:34lands the PR by itself. So you have a
- 8:37bunch of these bottlenecks that needs to
- 8:39be solved before you can transport the
- 8:41user entirely into into the cloud and
- 8:45really enable long running, multi hour
- 8:47running cloud agents. So that was over
- 8:50flex 1.0 which we launched about
- 8:53about a couple of months ago, I believe.
- 8:55And that has sort of like really taken
- 8:58off and we learned a lot from.
- 9:00[clears throat] So the biggest thing
- 9:01that we realized was now the bottleneck
- 9:04changed. Now that long running agents
- 9:06are possible, we were able to run a lot
- 9:09more tasks, you know, if desktop coding
- 9:11sort of enabled maybe three to five
- 9:14things in parallel, long running cloud
- 9:16agents started enabling 20 30 things in
- 9:18parallel, right? So, you know, obviously
- 9:20it sort of provided a lot more leverage
- 9:22and now the bottleneck sort of changed
- 9:25from writing code to project management
- 9:28and you know, making sure like the stuff
- 9:32lands and is actually useful
- 9:35um in making sure the things that are in
- 9:38my head are actually started.
- 9:40You know, because I have a limited
- 9:41context window, once it sort of like
- 9:43exceeds um 10 tasks in parallel, it gets
- 9:47like really really hard. So the
- 9:49bottleneck moves into project management
- 9:51actually.
- 9:52The model advances of the recent months
- 9:54also point in this direction. The latest
- 9:57gains have come from orchestration. So
- 10:01you know, most of the latest models try
- 10:03to offload
- 10:05um the chunk of heavy lifting into sub
- 10:08agents and the main agent, you know,
- 10:11wants to act as a orchestrator so that
- 10:14you maximally use
- 10:15the maximally generate intelligence
- 10:18within a context window. So fresh
- 10:20context windows
- 10:22give a lot more intelligence budget and
- 10:25you know, that's the way things are
- 10:26pointing.
- 10:27So that enables
- 10:30a really nice project manager which
- 10:32starts and orchestrates a bunch of
- 10:33sessions and
- 10:36and lands things for you. So that's
- 10:38WorxFlex 2.0. WorxFlex 2.0 is basically
- 10:42a bunch of autobots. These are 24/7
- 10:45employees that work towards a goal.
- 10:48The the bottleneck actually changed from
- 10:53giving a task and getting getting it
- 10:54done to actually even starting the task.
- 10:57The one thing that we started noticing
- 10:59was there was a lot of incremental
- 11:02intent that you could just capture from
- 11:04the way things were happening.
- 11:06For example, like if a customer call
- 11:10ends with a bunch of to-dos, those are
- 11:12all intent that I don't need to like
- 11:14copy paste and start one at a time.
- 11:17So [clears throat]
- 11:18like the mere act of even like noticing,
- 11:21remembering and starting each one of
- 11:23them could just be automated.
- 11:26[clears throat] So most teams treat
- 11:28agents like a task queue and a task
- 11:30queue only moves as fast to the person
- 11:32feeding it. But if you actually look at
- 11:33what's in the queue, most of it doesn't
- 11:36need you. The end to end plus one, that
- 11:39intent is already there in the code
- 11:41base. The alert or the transcript, it
- 11:43doesn't need a genius to start it. You
- 11:45can you can sort of figure out from the
- 11:47customer transcript, this is what you
- 11:49want to start. You can figure out from a
- 11:51production alert
- 11:53what you need to solve. Those things
- 11:55need to be on autopilot. It just needs
- 11:57an owner
- 11:59and it needs to be like multiplayer and
- 12:01collaborative with the rest of the team
- 12:03who where the ownership might be
- 12:04distributed to make sure all these sort
- 12:07of parallel sessions actually land. This
- 12:09is the next phase where of WorxFlex
- 12:12um and even um coding agents in general
- 12:17where things are moving from single
- 12:18player to like truly multiplayer and
- 12:20moving from um reactive to proactive
- 12:24where the intents that are already in
- 12:26there are being captured and used. So I
- 12:28would say like you know there are two
- 12:29kinds of work there is zero truly zero
- 12:31to one work that's like the truly
- 12:33creative stuff where the customer can't
- 12:35really even ask for it it takes taste or
- 12:37judgment that's the part that's still
- 12:39human
- 12:40and you know I I like doing that stuff
- 12:43and the part that's going into the auto
- 12:45bots is the end to end plus one stuff
- 12:48where the intents are sufficiently
- 12:50present as a part of the signals that
- 12:52the company is getting. So that's the
- 12:54new bottleneck and that's what Warflex
- 12:552.0 um is all about solving. Now people
- 12:59have tried to automate it and have had
- 13:01scars um doing something fairly simple.
- 13:05So for example you know people have set
- 13:07up automations to solve all sentries.
- 13:09You tell it to solve all the sentries
- 13:10and it goes crazy like a flood of PRs
- 13:13just hit you that nobody has times to
- 13:15time to merge. It pings you all day like
- 13:18one at a time and it's noisy. It often
- 13:21just silences the alert and leaves out
- 13:23the root cause um and a token bill that
- 13:25bursts um while the agent runs for hours
- 13:28and something nobody wanted. It's a
- 13:30black box you can't trust it with
- 13:31anything that matters right?
- 13:33You know these are all the fears that
- 13:36that people end up with when they just
- 13:39like point um a fire hose of incoming
- 13:43events company events on on an agent and
- 13:47or an automation. That doesn't really
- 13:49work. People have burnt their fingers.
- 13:52What you what
- 13:53what you need is a bot which actually
- 13:55has common sense. Um think about a new
- 13:57hire you don't fully trust yet. You give
- 13:59them important things first. You expect
- 14:02them to check in a lot early. You give
- 14:05them more as their work holds up and
- 14:07less when it doesn't. They work inside a
- 14:09budget and you can see what they do.
- 14:11Nobody hires a person and says solve
- 14:13everything.
- 14:14That's the bar that we built um for
- 14:17Warflex 2.0 and it's the next rung on
- 14:19the ladder. 1.0 finishes the work that
- 14:22you start.
- 14:23it. It gives you autobots, 24/7
- 14:26coworkers that live in You give an
- 14:27autobot a goal
- 14:29and a home channel. It stays plugged
- 14:31into your webhooks and starts the work
- 14:33itself.
- 14:35We even built a directory of opinionated
- 14:37bots, right? And you can you know start
- 14:40from there and change it. That's the big
- 14:42thing, you know people can't imagine
- 14:45these bots. So we've built a directory
- 14:47and we've sort of made sure
- 14:50the ones that we like and works for us,
- 14:53it's available out of the box with my
- 14:56taste and judgment to you.
- 14:59So so one [snorts] of some of the most
- 15:00you know I'll just walk you through some
- 15:02of the most well used bots. We have an
- 15:05on con on call bot that triages sentry
- 15:08and page duty alerts and investigates
- 15:10before an engineer is looped in. A CI
- 15:12bot that keeps speeding up your builds.
- 15:15A project manager bot which we started
- 15:18using a lot which holds a whole project
- 15:20in one channel and reports progress and
- 15:22blockers every day.
- 15:24This is super cool because it can it can
- 15:26make progress on things. It can tag
- 15:28people for reviews once
- 15:31some work needs to be reviewed. It can
- 15:33make your project run on
- 15:36on insane speed.
- 15:37One of the nice autobots that we have we
- 15:40have is like it it watches customer
- 15:42calls.
- 15:43For example, you know I was rolling out
- 15:46our beta to a friendly customer and the
- 15:48bot had a bug. It didn't respond back in
- 15:50the Slack channel right after. So you
- 15:52know normally I would have taken a
- 15:53screenshot and tagged Warflex
- 15:57to make it make it go work on the bug.
- 15:59This time I just you know after the call
- 16:01ended the autobot just read the
- 16:02transcript and found made to do located
- 16:05on the screen at that timestamp because
- 16:07you know it also has the you know video
- 16:09alongside the screen share alongside.
- 16:12And it it checked with me whether it
- 16:15wants me to fix it and I said yes. it
- 16:18tagged the right engineer who who don't
- 16:20the fix and you know came back with the
- 16:23PR. The engineer approved it. It it got
- 16:26merged.
- 16:28It got deployed. It closed the item and
- 16:30it even drafted a follow-up to the
- 16:32customer. I approved the follow-up and
- 16:36it went it went to the customer that it
- 16:38was fixed, right? So that that's the
- 16:40kind of leverage, you know, you just
- 16:41like I think the future is just like
- 16:43approve approve approve. A lot of these
- 16:44routine stuff is like you you just
- 16:46approve it and even that like a bot sort
- 16:48of like learns
- 16:51how much confidence it has and like
- 16:53keeps moving forward and over time it
- 16:54gets better and better like an employee.
- 16:56So
- 16:57[clears throat]
- 16:58one one thing for this room on why it
- 17:00works now, the latest model training
- 17:02runs focused on orchestration, a main
- 17:04agent that plans and keeps track with
- 17:06heavy lifting pushed down to sub agents.
- 17:08Most frontier models can now solve a
- 17:11well-defined task and their output is
- 17:13hard to tell apart. The next level up is
- 17:15a manager, an agent that starts sessions
- 17:17and keeps track of them. That's an
- 17:18autobot. It also runs on what you
- 17:20already have. Most teams bring their
- 17:23cloud code or codex subscriptions.
- 17:25Larger API users can route the well
- 17:28specified work to open models and keep
- 17:30the frontier models for judgment and
- 17:32communication. That way you can save a
- 17:33lot of money. Obviously the subscription
- 17:36save you a lot of money or if you're not
- 17:38eligible for it, the open source models
- 17:40save you a lot of money. What we built
- 17:42is an agent that decides what to work on
- 17:43next inside the limits you set and shows
- 17:46you why. Ultimately if you follow
- 17:48maximizing the leverage per person to
- 17:51its end, you get a company that runs on
- 17:54loops. A customer call ends and the
- 17:56product gets better. An alert fires and
- 17:59someone already knows why. CI slows down
- 18:02and a tested fix is waiting for review.
- 18:06The people spend their day on zero to
- 18:07one and on building the next loop. Our
- 18:10not star is a zero employee company. Our
- 18:13not star tells you which way to walk.
- 18:16It doesn't say that you arrive. 10
- 18:18people driving a huge business rounds
- 18:20down to
- 18:21We're early.
- 18:21Autobots will get things wrong. Some
- 18:23responsibilities are not ready to hand
- 18:25over,
- 18:26but we're building the next LEGO block,
- 18:29you know, the larger LEGO block which is
- 18:32assembled with the underlying model.
- 18:34Tesla took ground one intervention at a
- 18:36time where the way they solved autopilot
- 18:40is by profiling
- 18:42which
- 18:43which areas or scenarios the model made
- 18:46mistake. And that's what we do. We
- 18:48profile where humans spend their time
- 18:51and what the bottlenecks are and figure
- 18:53out how to throw tokens and intelligence
- 18:56at the problem. We climb the leverage
- 18:59one rung at a time as a new bottleneck
- 19:00shows up. We
- 19:02try to We are probably in a super
- 19:04intelligence takeoff and the
- 19:07opportunities to rewrite much of how the
- 19:09world runs are like insanely high.
- 19:11Companies build record revenue
- 19:14in super small amount of time. The
- 19:16growth rates are incredible, something
- 19:19that I've never seen all my life. But
- 19:20also it's like one of the most
- 19:22competitive spaces ever.
- 19:26The teams that tend to win tend to be
- 19:29like really small with
- 19:31everyone having a lot of ownership and
- 19:33leverage. It's It's not the traditional
- 19:35incumbents are not able to adapt and
- 19:37like
- 19:39and compete.
- 19:40There's a land grab and it goes to small
- 19:43teams with taste, judgment, and an
- 19:46insane amount of leverage per person.
- 19:48Own the zero to one and take as much
- 19:50ground as you can in this land grab.
- 19:52Thank you so much, guys. Really
- 19:54appreciate the opportunity speaking with
- 19:56all of you guys today. Thank you.
About this transcript
This page contains the full transcript of Merge Is the Bottleneck - Prasanna Sankar, Vorflux by AI Engineer, generated from the public captions YouTube serves with the video. The transcript has 3,463 words across 542 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.