Loading video...

Video Failed to Load

There was a problem loading this video. This could be due to a temporary network issue or the video might be unavailable.

SpaceXAI engineer Lingxi Li, Lingxi Li, runs code cleanup while everyone else is asleep. Nobody is shipping at 3 AM, so conflicts are rare and the changes are low risk. His bots start a research agent that sweeps the monorepo for quality issues, poor modularization, comment bloat and security... show more

tetsuo

241,178 subscribers

1,002,200 views • 14 days ago •via X (Twitter)

Science & Technology

Anya Rossi• Live Now

Private livecam show

22 Comments

Arthur MacWaters14 days ago

@lingxi @lingxi is the goat, great guy also

Mike Melgarejo14 days ago

@lingxi a special mind

tetsuo14 days ago

@lingxi Facts.

Merdog14 days ago

@lingxi So dis bro is da one shipping in the middle of my grinds Bahahha

John Bai14 days ago

@lingxi bro is agi

arclight industries14 days ago

@lingxi Real G's move in silence

Evan Bacon 🥓14 days ago

@lingxi Lingxi is an absolute pro 💪

decipherx14 days ago

@lingxi give every overnight cleanup a kill switch. if tests miss a behavior change, the first engineer online should be able to revert the whole batch before reading the summary

Devin Soto14 days ago

@lingxi This is actually the best way to ship

why14 days ago

@lingxi The low-conflict window is interesting, but the real unlock is having the agent surface small, reviewable changes instead of opening one giant cleanup PR.

Vito14 days ago

@elonmusk @lingxi That guy is brilliant

K ι T-7Y Wolf❤️🫶❤️14 days ago

@lingxi 3am club 😎🫡

Alek14 days ago

@lingxi 3 AM merges with zero conflicts?

Aditya Kumar14 days ago

@lingxi oh wow, we used to do that at @Meta and nobody even noticed. Like upgraded C++ version of a very large codebase with only a handful of people. But AI does it and suddenly it is cool!

The Retarded Elon14 days ago

@lingxi they talked about peadpod too :)

Eternities AI14 days ago

@lingxi 3am is cheap because conflict rate drops, not because judgment does. overnight cleanup only works if the bot can fail closed and leave a receipt a morning human can reject without archaeology.

Tree Lake Rain 🌲🦆🌦️14 days ago

@lingxi

Hey Its Burt14 days ago

@lingxi Facts: When you work with a lot of people, sometimes find times when you can work off-hours from everybody else. Or have the bots work when nobody else is working and coordinate with the other bots who happen to be working.

Sam Liu14 days ago

@lingxi Overnight agents are compelling because they turn idle hours into a durable inbox, not because they merge unattended. In @puffoai, agents can work in parallel, leave results in shared threads, and hand consequential changes back to humans for review.

Daniel Silva14 days ago

@lingxi Ohh

Mar Jo14 days ago

@elonmusk @lingxi 👏👏👏👏👏

Daniel Silva14 days ago

@lingxi Great 👍🏻 follow me

Related Videos

SpaceXAI engineer, Lingxi Li: "My team of 10+ GrokBot agents is shipping code at 3 am, when all SpaceXAI team is sleeping, conflicts are rare and the changes are low risk. my Chief of Engineers agent rules PM and 10+ coding agents in a graph. My job is to build the right team" In a 30-minute workshop, a SpaceXAI engineer revealed his engineering setup for GrokBot worth more than a $500 agentic engineering course watch today, then read article below on how to build a team of GrokBot agents that will work for you 24/7

SpaceXAI engineer, Lingxi Li: "My team of 10+ GrokBot agents is shipping code at 3 am, when all SpaceXAI team is sleeping, conflicts are rare and the changes are low risk. my Chief of Engineers agent rules PM and 10+ coding agents in a graph. My job is to build the right team" In a 30-minute workshop, a SpaceXAI engineer revealed his engineering setup for GrokBot worth more than a $500 agentic engineering course watch today, then read article below on how to build a team of GrokBot agents that will work for you 24/7

Codez

13,370 views • 13 days ago

SpaceXAI just dropped a GrokBot that works for you while you sleep: "I am running this GrokBot before I go to sleep. I wake up and task is done" here's how the Nightly Engineer works: • every night → researches the whole codebase and finds cleanup opportunities • one area → one focused cleanup PR instead of one giant refactor • verification → reproduce the issue first, then test the actual fix • by 4am → the audit is done and your cleanup PRs are waiting the result: → you go to sleep with a messy codebase → you wake up with researched, scoped cleanup PRs SpaceXAI basically turned code maintenance into a 24/7 GrokBot routine save now, then read how to build a 24/7 GrokBot agent team in the article below

SpaceXAI just dropped a GrokBot that works for you while you sleep: "I am running this GrokBot before I go to sleep. I wake up and task is done" here's how the Nightly Engineer works: • every night → researches the whole codebase and finds cleanup opportunities • one area → one focused cleanup PR instead of one giant refactor • verification → reproduce the issue first, then test the actual fix • by 4am → the audit is done and your cleanup PRs are waiting the result: → you go to sleep with a messy codebase → you wake up with researched, scoped cleanup PRs SpaceXAI basically turned code maintenance into a 24/7 GrokBot routine save now, then read how to build a 24/7 GrokBot agent team in the article below

0xRafy

16,491 views • 11 days ago

SpaceXAI engineer Lauren Tan: "I'm running 20+ GrokBot agents with /loop and /goal inside pstack My bots ship code while I sleep At SpaceXAI, 85% of engineers are already running 10+ GrokBot agents as a team One Chief of Staff agent knows every other bot and manages the entire system" GrokBot → /loop → /goal → Chief of Staff → Agent Teams In a 1-hour workshop, Lauren shows how to build a team of GrokBot agents from scratch Research → Code → Review → Ship Worth more than most $1000 agent engineering courses Bookmark and watch it today Then read the full article on building a GrokBot agent team from scratch

SpaceXAI engineer Lauren Tan: "I'm running 20+ GrokBot agents with /loop and /goal inside pstack My bots ship code while I sleep At SpaceXAI, 85% of engineers are already running 10+ GrokBot agents as a team One Chief of Staff agent knows every other bot and manages the entire system" GrokBot → /loop → /goal → Chief of Staff → Agent Teams In a 1-hour workshop, Lauren shows how to build a team of GrokBot agents from scratch Research → Code → Review → Ship Worth more than most $1000 agent engineering courses Bookmark and watch it today Then read the full article on building a GrokBot agent team from scratch

kaori

100,589 views • 26 days ago

Simps are the worst. - A girl hits a toto driver with her scooty. - A simp comes to play the knight. - He slaps that poor boy. - The poor boy starts crying. - There are so many men there. - Nobody supported that poor boy. He is working at such a young age to take care of his family, including his mother and sisters. All bootlicker simps need to go.

Simps are the worst. - A girl hits a toto driver with her scooty. - A simp comes to play the knight. - He slaps that poor boy. - The poor boy starts crying. - There are so many men there. - Nobody supported that poor boy. He is working at such a young age to take care of his family, including his mother and sisters. All bootlicker simps need to go.

︎ ︎venom

868,551 views • 2 years ago

❗️Cars in Russia are breaking down due to poor-quality gasoline, with drivers reporting a surge in Check Engine warnings The issue is linked to low-grade Euro-3 fuel, which is appearing at more and more gas stations across the country. Owners of Chinese cars are especially at risk, as their turbocharged engines and direct fuel injection systems are particularly sensitive to low-quality fuel. 🚲 At this rate, a bicycle might soon become the most reliable mode of transport in Russia.

❗️Cars in Russia are breaking down due to poor-quality gasoline, with drivers reporting a surge in Check Engine warnings The issue is linked to low-grade Euro-3 fuel, which is appearing at more and more gas stations across the country. Owners of Chinese cars are especially at risk, as their turbocharged engines and direct fuel injection systems are particularly sensitive to low-quality fuel. 🚲 At this rate, a bicycle might soon become the most reliable mode of transport in Russia.

NEXTA

770,515 views • 3 months ago

At a product launch event last year, Li Bin, ceo of NIO, said “those who think of LiDARs being useless are either dumb or evil. LiDARs can make up for the weakness in vision-only perception”, followed by a short clip showing that LiDAR is able to recognize the obstacle (a stack of tires) in pitch dark and come to a full stop. That was his comment made before FSD rolled out in China. Pretty sure he has watched plenty of FSD videos and wonder what he has to say now.

At a product launch event last year, Li Bin, ceo of NIO, said “those who think of LiDARs being useless are either dumb or evil. LiDARs can make up for the weakness in vision-only perception”, followed by a short clip showing that LiDAR is able to recognize the obstacle (a stack of tires) in pitch dark and come to a full stop. That was his comment made before FSD rolled out in China. Pretty sure he has watched plenty of FSD videos and wonder what he has to say now.

Ray

70,324 views • 1 year ago

𝐓𝐘𝐋𝐄𝐑𝟏 𝐌𝐄𝐄𝐓𝐒 𝐒𝐎𝐍 𝐎𝐅 𝐀𝐑𝐔𝐆𝐀𝐋 𝐈𝐍 𝐇𝐀𝐑𝐃𝐂𝐎𝐑𝐄 𝐂𝐋𝐀𝐒𝐒𝐈𝐂 Son of Arugal is a rare in Silverpine Forest who is known to decimate low levels that are unfortunate enough to come across him. loltyler1 experiences this first hand when he stumbles across the rare and watches a fleeing lowbie die before his eyes. Tyler1 is quick to recognize the danger and runs away.

Sensitive content

𝐓𝐘𝐋𝐄𝐑𝟏 𝐌𝐄𝐄𝐓𝐒 𝐒𝐎𝐍 𝐎𝐅 𝐀𝐑𝐔𝐆𝐀𝐋 𝐈𝐍 𝐇𝐀𝐑𝐃𝐂𝐎𝐑𝐄 𝐂𝐋𝐀𝐒𝐒𝐈𝐂 Son of Arugal is a rare in Silverpine Forest who is known to decimate low levels that are unfortunate enough to come across him. loltyler1 experiences this first hand when he stumbles across the rare and watches a fleeing lowbie die before his eyes. Tyler1 is quick to recognize the danger and runs away.

N_Tys

363,510 views • 1 year ago

The SpaceXAI team just leaked the Grok Bots they run on themselves. Ten installs. A $1,000,000/year org chart if you hired this as people. 1. Fondi, Nao (Naoufal), Field Engineer at SpaceXAI Reads your public site and stands up a leadership suite. Installing Grok Bot plus this bot is the hire. 2. loops, Matt Palmer, Cursor + Grok Bot at SpaceXAI The outer loop he runs engineering on. Sits above coding agents, writes the goal, reviews, merges. You name the repo. 3. Projects Manager, Eric Zakariasson, SpaceXAI (prev Cursor) One bot, then it spins up a coder, designer, researcher, writer. Notion is source of truth. Replaces a PM plus a four person pod. 4. Chieeeeefy, Nao (Naoufal), Field Engineer at SpaceXAI Chief of staff for a field engineer. Calendar and inbox first. The $120k EA seat. 5. Rutin, Nao (Naoufal), Field Engineer at SpaceXAI Monday tune up for every routine across the fleet. The ops hire that stops silent failures. 6. dr eggbot, Lauren (poteto), Grok Bot at SpaceXAI (prev Cursor) Builds other Grok Bots for you. Skip the consultant who charges five figures to set your agents up. 7. loom, Lauren (poteto), Grok Bot at SpaceXAI Reads your Gmail threads and drafts the reply. Never sends. An $80k EA inbox, minus the send button. 8. Querie, Emily Gavrilenko, building SpaceXAI (prev Cursor, former founder) Her data scientist bot. Owns every analytics query and spreadsheet pull so nobody else on the team touches them. 9. Echo, Krista Letz, Enterprise at SpaceXAI (prev Cursor) Turns a sales call into slides straight from the call notes. Runs automatically after every customer call, or mid call with /echo. 10. SE call bot, Scott Metcalf, GTM Productivity at SpaceXAI Live backup for solutions engineers on customer calls. The $180k SE sitting in the room. Save this. These are the people who ship the product 👇

The SpaceXAI team just leaked the Grok Bots they run on themselves. Ten installs. A $1,000,000/year org chart if you hired this as people. 1. Fondi, Nao (Naoufal), Field Engineer at SpaceXAI Reads your public site and stands up a leadership suite. Installing Grok Bot plus this bot is the hire. 2. loops, Matt Palmer, Cursor + Grok Bot at SpaceXAI The outer loop he runs engineering on. Sits above coding agents, writes the goal, reviews, merges. You name the repo. 3. Projects Manager, Eric Zakariasson, SpaceXAI (prev Cursor) One bot, then it spins up a coder, designer, researcher, writer. Notion is source of truth. Replaces a PM plus a four person pod. 4. Chieeeeefy, Nao (Naoufal), Field Engineer at SpaceXAI Chief of staff for a field engineer. Calendar and inbox first. The $120k EA seat. 5. Rutin, Nao (Naoufal), Field Engineer at SpaceXAI Monday tune up for every routine across the fleet. The ops hire that stops silent failures. 6. dr eggbot, Lauren (poteto), Grok Bot at SpaceXAI (prev Cursor) Builds other Grok Bots for you. Skip the consultant who charges five figures to set your agents up. 7. loom, Lauren (poteto), Grok Bot at SpaceXAI Reads your Gmail threads and drafts the reply. Never sends. An $80k EA inbox, minus the send button. 8. Querie, Emily Gavrilenko, building SpaceXAI (prev Cursor, former founder) Her data scientist bot. Owns every analytics query and spreadsheet pull so nobody else on the team touches them. 9. Echo, Krista Letz, Enterprise at SpaceXAI (prev Cursor) Turns a sales call into slides straight from the call notes. Runs automatically after every customer call, or mid call with /echo. 10. SE call bot, Scott Metcalf, GTM Productivity at SpaceXAI Live backup for solutions engineers on customer calls. The $180k SE sitting in the room. Save this. These are the people who ship the product 👇

unicode

48,600 views • 27 days ago

SpaceXAI engineer: "one month ago GrokBot was nothing. now it is the best agent product on market my Grok Bot agent built a product and hired 100+ agents, that are working on this project together" in a 30-minutes workshop, a SpaceXAI engineer shows how to make agent to call your customers Worth more than a 3-hours lecture on building agents watch today, then read how to build a chain of Grok agents from scratch below

SpaceXAI engineer: "one month ago GrokBot was nothing. now it is the best agent product on market my Grok Bot agent built a product and hired 100+ agents, that are working on this project together" in a 30-minutes workshop, a SpaceXAI engineer shows how to make agent to call your customers Worth more than a 3-hours lecture on building agents watch today, then read how to build a chain of Grok agents from scratch below

0xRafy

11,976 views • 18 days ago

Day 4 of building GTA 6 with a loop of AI agents. where it's at: - The character's got a model. so does every NPC. - Cars have models now - They're starting to move. actual driving - Traffic rules are showing up. - Performance jumped and there's a stack of pull requests I haven't merged yet. Agent is still looping. PRs are still open.

Day 4 of building GTA 6 with a loop of AI agents. where it's at: - The character's got a model. so does every NPC. - Cars have models now - They're starting to move. actual driving - Traffic rules are showing up. - Performance jumped and there's a stack of pull requests I haven't merged yet. Agent is still looping. PRs are still open.

Ziwen

60,284 views • 3 months ago

"We're trying to hire as many new grads as we can. We're finding that a lot of young people are agentmaxxing." - Replit founder Amjad Masad Amjad says that despite the “doom and gloom” around entry-level jobs, this is actually a “golden era” for young people who know how to use AI tools: "The roles are collapsing. We have designers shipping code, engineers shipping design, and salespeople shipping code." "The particular skill is not the bottleneck anymore. It is how ambitious you are, how generative you are, how creative you are, and how good you are at utilizing these tools." "So I think it's really a golden era for a lot of young people coming up right now in the job market. In our time, you would go and interview and become a Java J2EE software engineer doing global migrations. That was your title, right? It's no longer the case. We just want people to come and do things, and I think that's incredibly exciting."

"We're trying to hire as many new grads as we can. We're finding that a lot of young people are agentmaxxing." - Replit founder Amjad Masad Amjad says that despite the “doom and gloom” around entry-level jobs, this is actually a “golden era” for young people who know how to use AI tools: "The roles are collapsing. We have designers shipping code, engineers shipping design, and salespeople shipping code." "The particular skill is not the bottleneck anymore. It is how ambitious you are, how generative you are, how creative you are, and how good you are at utilizing these tools." "So I think it's really a golden era for a lot of young people coming up right now in the job market. In our time, you would go and interview and become a Java J2EE software engineer doing global migrations. That was your title, right? It's no longer the case. We just want people to come and do things, and I think that's incredibly exciting."

TBPN

47,144 views • 6 months ago

SpaceXAI engineer, ex-Cursor: "i haven't hired a human in a year, i just spin up another bot 50 of them run on a $200 plan, and they clear what used to be a 3-week sprint before i finish breakfast we did $1.2M this year on a headcount of one" in a 54-minute workshop, a SpaceXAI engineer shows the setup that runs 50 bots off one plan this is the piece almost nobody builds, and it's why their agents fall apart past three watch it, then build the same memory into your team with the 12-step system in the article below

SpaceXAI engineer, ex-Cursor: "i haven't hired a human in a year, i just spin up another bot 50 of them run on a $200 plan, and they clear what used to be a 3-week sprint before i finish breakfast we did $1.2M this year on a headcount of one" in a 54-minute workshop, a SpaceXAI engineer shows the setup that runs 50 bots off one plan this is the piece almost nobody builds, and it's why their agents fall apart past three watch it, then build the same memory into your team with the 12-step system in the article below

KingWilliam

1,385,068 views • 18 days ago

President William Ruto has revealed that he is setting up a research laboratory at his farm, saying innovation and scientific research should be central to Kenya’s development. Speaking at State House, Nairobi, where he awarded charters to KAIST and KEMRI, Ruto said equipment for the lab has already been procured and is on the way. Ruto: I am already developing a lab on my farm. I have gotten some equipment which is on its way for a lab at my farm because I believe the future of a nation is in making sure that research becomes front and centre of everything that we are doing.

President William Ruto has revealed that he is setting up a research laboratory at his farm, saying innovation and scientific research should be central to Kenya’s development. Speaking at State House, Nairobi, where he awarded charters to KAIST and KEMRI, Ruto said equipment for the lab has already been procured and is on the way. Ruto: I am already developing a lab on my farm. I have gotten some equipment which is on its way for a lab at my farm because I believe the future of a nation is in making sure that research becomes front and centre of everything that we are doing.

The Eastleigh Voice

68,393 views • 4 months ago

Mike Johnson admitting on a hot mic and camera that he is excited at the prospect of low voter turnout. This isn’t a joke. We all have to wake up right now people. They are trying to install a dictator. Everyone else in the world can see it except for this 1% of its population.

Mike Johnson admitting on a hot mic and camera that he is excited at the prospect of low voter turnout. This isn’t a joke. We all have to wake up right now people. They are trying to install a dictator. Everyone else in the world can see it except for this 1% of its population.

MAGA Cult Slayer

468,042 views • 1 year ago

Imagine a player who is so good that at the age of 29, he has 3 FIFA world cups in his name. They called him ‘O Rei’—King. Long before Messi and Ronaldo, Pelé was the real mistake in the system. He is a 5’8” powerhouse who can jump over the central defender, the winger runs faster and more skilled than everyone else. He has no sports science or modern throwing—he only has a ball and a dream that changes the sport forever. 1,281 goals, 3 World Cups and a legacy that will never be touched.

Imagine a player who is so good that at the age of 29, he has 3 FIFA world cups in his name. They called him ‘O Rei’—King. Long before Messi and Ronaldo, Pelé was the real mistake in the system. He is a 5’8” powerhouse who can jump over the central defender, the winger runs faster and more skilled than everyone else. He has no sports science or modern throwing—he only has a ball and a dream that changes the sport forever. 1,281 goals, 3 World Cups and a legacy that will never be touched.

Baskararayyan

78,819 views • 6 months ago

Musk suggested that his life could be at risk when he said that he needs to beef up his security. This is because “people who are taking advantage of the government” are going to get “upset” when he and Trump start slashing costs. “If I had full power to take action, I would balance the budget immediately,” Musk said, adding that the government urgently needs to stop spending more than it brings in. “A lot of people who are taking advantage of the government are going to be upset about that, and I’d probably need a lot of security. But it’s got to be done. And if it’s not done, we’ll just go bankrupt.”

Musk suggested that his life could be at risk when he said that he needs to beef up his security. This is because “people who are taking advantage of the government” are going to get “upset” when he and Trump start slashing costs. “If I had full power to take action, I would balance the budget immediately,” Musk said, adding that the government urgently needs to stop spending more than it brings in. “A lot of people who are taking advantage of the government are going to be upset about that, and I’d probably need a lot of security. But it’s got to be done. And if it’s not done, we’ll just go bankrupt.”

The Vigilant Fox 🦊

18,339,564 views • 1 year ago

Tommy Robinson is laughing at his gullible supporters while he and his son are living a high luxury life. If he is a priority target for ISIS, why is he keep updating everyone about the next location he's going to be at.

Tommy Robinson is laughing at his gullible supporters while he and his son are living a high luxury life. If he is a priority target for ISIS, why is he keep updating everyone about the next location he's going to be at.

Mukhtar

56,786 views • 7 months ago

acpx v0.4 ships Agentic Workflows, or as I like to call them "Agentic Graphs" It let's you create node-based workflows on top of ACP (Agent Client Protocol), to drive any coding agent (Codex, Claude Code, pi) through deterministic steps This let's you automate routine, mechanical legwork like triaging incoming PRs, bugs in error reporting, and so on... For example, OpenClaw receives 300~500 new PRs per day. A lot of them are low quality, but they still relate to real issues, so you have to address them somehow You need to: - extract the intent - cluster them based on intent - figure out if the proposed changes are legit, or whether they are slop local solutions, like trying to catch flies instead of drying out the swamp - if the PR is too low quality or the intent is not clear, close them - run AI review on them them and address any issues that come up - refactor them if the changes are half-baked - resolve conflicts - and so on... So that when the PR is presented to the attention of the maintainer, all the routine legwork is done and the only remaining thing is the decision to (a) merge, (b) give feedback to the PR author, or (c) take over the PR work yourself I wanted to build this feature since a couple months now, since Codex got so good. OpenAI models are now good at judging implementation quality, so I found myself repeating the same steps I wrote above over and over I also tried putting all this in a single prompt. But I believe there are workflows that should not be a single prompt, but a sequence of prompts in the same session That is because like humans, LLMs are prone to PRIMING. I claim that putting all steps in the same prompt at the beginning of the context will generally give suboptimal results, compared to revealing the intention to the model step by step Creating such a workflow also gives more OBSERVABILITY into the each step that an agent is supposed to take. Agent generates JSON at the end of each step, and that structured data can be used to monitor thousands of agents running at the same time in an easier way, on a dashboard Similar features have been introduced in e.g. n8n, langflow. But AFAIK they are not integrating ACP like the way I do I wanted to have a fresh approach, and to build an API that I can develop freely the way I want, so I created a new workflow API inside acpx The video is from the workflow run viewer, but that is not where you build the workflow. You build it by using the acpx flow typescript API. See examples/pr-triage in acpx repo Before building that, I started from a Markdown file with a Mermaid chart of the flow I had in mind. The Markdown file acts as a spec for the flow, and I have built the workflow through trial and error. I call this process "workflow tuning" I started working on acpx repo PRs one by one, tuning the flow, slowly scaling to more PRs. Finally, when I felt confident, I ran it in parallel over all external open PRs in the acpx repo. I believe it already saved me hours this week My next goal, if well received, is to set this up on a cloud agent so that it can process the 300~500 PRs the OpenClaw repo receives every day, in real time, as they come in I believe this will save all open source maintainers around the world countless hours and make it much easier to herd and absorb external contributions from everyone!

acpx v0.4 ships Agentic Workflows, or as I like to call them "Agentic Graphs" It let's you create node-based workflows on top of ACP (Agent Client Protocol), to drive any coding agent (Codex, Claude Code, pi) through deterministic steps This let's you automate routine, mechanical legwork like triaging incoming PRs, bugs in error reporting, and so on... For example, OpenClaw receives 300~500 new PRs per day. A lot of them are low quality, but they still relate to real issues, so you have to address them somehow You need to: - extract the intent - cluster them based on intent - figure out if the proposed changes are legit, or whether they are slop local solutions, like trying to catch flies instead of drying out the swamp - if the PR is too low quality or the intent is not clear, close them - run AI review on them them and address any issues that come up - refactor them if the changes are half-baked - resolve conflicts - and so on... So that when the PR is presented to the attention of the maintainer, all the routine legwork is done and the only remaining thing is the decision to (a) merge, (b) give feedback to the PR author, or (c) take over the PR work yourself I wanted to build this feature since a couple months now, since Codex got so good. OpenAI models are now good at judging implementation quality, so I found myself repeating the same steps I wrote above over and over I also tried putting all this in a single prompt. But I believe there are workflows that should not be a single prompt, but a sequence of prompts in the same session That is because like humans, LLMs are prone to PRIMING. I claim that putting all steps in the same prompt at the beginning of the context will generally give suboptimal results, compared to revealing the intention to the model step by step Creating such a workflow also gives more OBSERVABILITY into the each step that an agent is supposed to take. Agent generates JSON at the end of each step, and that structured data can be used to monitor thousands of agents running at the same time in an easier way, on a dashboard Similar features have been introduced in e.g. n8n, langflow. But AFAIK they are not integrating ACP like the way I do I wanted to have a fresh approach, and to build an API that I can develop freely the way I want, so I created a new workflow API inside acpx The video is from the workflow run viewer, but that is not where you build the workflow. You build it by using the acpx flow typescript API. See examples/pr-triage in acpx repo Before building that, I started from a Markdown file with a Mermaid chart of the flow I had in mind. The Markdown file acts as a spec for the flow, and I have built the workflow through trial and error. I call this process "workflow tuning" I started working on acpx repo PRs one by one, tuning the flow, slowly scaling to more PRs. Finally, when I felt confident, I ran it in parallel over all external open PRs in the acpx repo. I believe it already saved me hours this week My next goal, if well received, is to set this up on a cloud agent so that it can process the 300~500 PRs the OpenClaw repo receives every day, in real time, as they come in I believe this will save all open source maintainers around the world countless hours and make it much easier to herd and absorb external contributions from everyone!

Onur Solmaz

149,693 views • 6 months ago