Lauren Tan (poteto) on the trust ladder, verification and shipping 2,500 PRs a month: analysis

Transcript · Report with the audio

Matt Pocock interviews Lauren Tan (poteto) in a Q&A about her talk on shipping 2,500 PRs to production in a month (Matt Pocock 1:01). She describes how she came to trust agents with more work step by step, which she calls a “trust ladder”. It started with a side project after she left Meta and continued with performance work at Cursor (Lauren Tan (poteto) 1:58; Lauren Tan (poteto) 5:08). Her main claims are that verification is the most important skill and that an engineer’s new job is to build a constrained environment where bad code is hard to write (Lauren Tan (poteto) 16:14; Lauren Tan (poteto) 26:14; Lauren Tan (poteto) 30:31). To scale, she runs personal agents that watch Slack, X and Linear and pass work to Cursor Projects coordinator agents (Lauren Tan (poteto) 42:09). Agents now merge their own PRs overnight (Lauren Tan (poteto) 51:57), and she reviews by sampling (Lauren Tan (poteto) 49:31). They ended on how to combine their skill libraries and agreed to talk again (Lauren Tan (poteto) 1:01:26; Lauren Tan (poteto) 1:05:16).

People

• Lauren Tan (poteto) (79.0% of talking time, 9,495 words)
• Matt Pocock (21.0% of talking time, 2,780 words): Host, likely Matt Pocock Pocock. Lauren credits him with Total TypeScript (Lauren Tan (poteto) 28:44), he has a popular skills library (Matt Pocock 0:47), and he is filming a course on software factories (Matt Pocock 10:31; Matt Pocock 49:04). He ran the conversation as a Q&A on her talk and steered it toward practical questions: triggers, her day-to-day setup, review and one-way doors (Matt Pocock 33:02; Matt Pocock 40:23; Matt Pocock 55:53). He questioned calling her setup a “dark factory” and suggested a dimmer switch instead (Matt Pocock 54:49; Matt Pocock 55:12).

Chapters

1. Introduction [0:00]. Matt Pocock introduces poteto as the guest after an earlier episode with Uncle Bob (Matt Pocock 0:00). He jokes about two popular skill libraries meeting (Matt Pocock 0:47) and says he will run the conversation as a Q&A on her viral talk about shipping 2,500 PRs in a month (Matt Pocock 1:01).
2. How the trust ladder started [1:48]. After Meta she took a month off for burnout and started a side project. She found herself micromanaging one agent and began writing skills to pass her own way of working to the agent. That early skill set, poteto/noodle, became the basis of pstack (Lauren Tan (poteto) 1:58; Lauren Tan (poteto) 3:32). She joined Cursor in March and worked on the performance of the agents window, where she found she was the “meat proxy” between her agent and Chrome DevTools (Lauren Tan (poteto) 4:20; Lauren Tan (poteto) 5:08; Lauren Tan (poteto) 5:22). Because agents take shortcuts, her skills aim to make the easy thing the right thing (Lauren Tan (poteto) 5:22; Lauren Tan (poteto) 6:09).
3. Domain expertise and language [6:22]. She argues domain expertise matters more than ever, because the bottleneck is now how clearly you express intent to the agent (Lauren Tan (poteto) 6:50; Lauren Tan (poteto) 7:38). Matt Pocock describes picking words that change how agents think, with TDD and grilling as examples (Matt Pocock 8:48). She cites his tip on tautological tests as a case of a lot of meaning packed into one word (Lauren Tan (poteto) 9:23).
4. Michelin kitchen instead of software factory [10:31]. Matt Pocock mentions his upcoming course on software factories and asks about her alternative metaphor (Matt Pocock 10:31; Matt Pocock 10:47). She finds “factory” apt but says it does not suggest quality or craft, and a Michelin kitchen is more aspirational. The kitchen also matches the trust ladder: a home cook alone, then a kitchen with others in it, then a chef who acts as tech lead and organizes everything (Lauren Tan (poteto) 11:18; Lauren Tan (poteto) 12:50; Lauren Tan (poteto) 13:30). In her view the environment, skills and codebase are the new ingredients (Lauren Tan (poteto) 14:17).
5. Verification and deterministic CLIs [15:12]. Matt Pocock praises her focus on the agent’s environment and on verification (Matt Pocock 15:12). She calls verification the most important skill because it lets the agent run and inspect the code itself. Verification is what makes a loop a loop, and it allows hill climbing, where an agent keeps improving against a score (Lauren Tan (poteto) 16:14; Lauren Tan (poteto) 17:04; Lauren Tan (poteto) 18:06). Every Cursor app now has an auto-maintained verification skill (Lauren Tan (poteto) 18:53). Her verification CLI wraps Playwright and the Chrome DevTools Protocol. It takes deterministic steps out of agent judgment and stops each agent from rebuilding the same tooling (Lauren Tan (poteto) 20:32; Lauren Tan (poteto) 22:09; Lauren Tan (poteto) 22:47). She applies the same principle to migrations, using codemods (Lauren Tan (poteto) 23:39; Lauren Tan (poteto) 24:33).
6. Environment, constraints and the Dune framework [24:52]. Matt Pocock defines a good codebase as one that is easy to change, with guardrails that keep agents and humans on narrow paths (Matt Pocock 25:06; Matt Pocock 25:32). She says the engineer’s new job is to work on the environment, and compares an engineer who skips this to a cook with dull knives (Lauren Tan (poteto) 26:14; Lauren Tan (poteto) 27:05). She links constraints to type narrowing in TypeScript. She describes Dune, an internal framework for Cursor’s Electron apps that works like an internal Next.js, with strict lint rules and one directory per feature (Lauren Tan (poteto) 28:44; Lauren Tan (poteto) 29:38; Lauren Tan (poteto) 30:31). She turns repeated agent failures into lint rules, and she broke up early GroqBot versions that consisted of huge God files (Lauren Tan (poteto) 31:16).
7. Scaling: inner and outer loops [32:55]. Matt Pocock asks where 2,500 PRs come from (Matt Pocock 33:02). She says the environment is the prerequisite, and that each big project is a restaurant she moves between (Lauren Tan (poteto) 33:32; Lauren Tan (poteto) 34:17). Her inner loop is agents writing code toward a snapshot of her intent. Her outer loop brings in outside context such as bug reports from Slack, Linear and X (Lauren Tan (poteto) 35:08; Lauren Tan (poteto) 35:58). Connecting the two lets agents triage and reproduce bugs themselves. Her guiding question is where she is the bottleneck (Lauren Tan (poteto) 36:46; Lauren Tan (poteto) 37:43; Lauren Tan (poteto) 38:33).
8. Concrete setup: GroqBot and Cursor Projects [40:10]. Matt Pocock asks what her setup looks like day to day (Matt Pocock 40:23). She runs GroqBot personal agents with connectors, which have routines that watch Slack, X, email and Linear. She ties this to the Netflix principle “context, not control” (Lauren Tan (poteto) 40:49; Lauren Tan (poteto) 41:46; Lauren Tan (poteto) 42:09). The GroqBots send work to Cursor Projects, coordinator agents in the cloud that delegate to sub-agents. A coordinator can group related bug reports to find a shared root cause (Lauren Tan (poteto) 42:56; Lauren Tan (poteto) 43:34; Lauren Tan (poteto) 44:22).
9. Gardening and buffered routines [45:28]. She says many of the PRs are gardening, not features, and that a good environment also helps human new hires (Lauren Tan (poteto) 45:51; Lauren Tan (poteto) 46:43). One routine looks for React footguns and adds them to a document instead of fixing them. She reads the document every few days to spot patterns she would miss fixing bugs one at a time (Lauren Tan (poteto) 47:17; Lauren Tan (poteto) 48:06). Matt Pocock says this may change the course he films next week (Matt Pocock 49:04).
10. Review, sampling and the dark factory [49:15]. She reviews by sampling, like a quality supervisor in a factory. When several agents repeat a mistake, she fixes the environment, not the single agent (Lauren Tan (poteto) 49:31; Lauren Tan (poteto) 50:27). She says her setup is in fact a dark factory. More than 10 coordinator agents work around the clock, and pstack’s “full autopilot” mode runs verifier agents that fuzz each PR before it merges (Lauren Tan (poteto) 51:57; Lauren Tan (poteto) 52:49). She reviews the commit history in the morning (Lauren Tan (poteto) 53:37; Lauren Tan (poteto) 53:59). Matt Pocock argues it is more like a dimmer switch, and she answers that the restaurant metaphor fits better (Matt Pocock 55:12; Lauren Tan (poteto) 55:23; Lauren Tan (poteto) 55:27).
11. One-way doors and verifiable domains [55:44]. Matt Pocock asks about fields such as medicine, law and finance, where a merged PR may be a one-way door that cannot be reverted (Matt Pocock 55:53; Matt Pocock 56:16). She says it depends on how verifiable the domain is, and admits she has no full answer (Lauren Tan (poteto) 56:28; Lauren Tan (poteto) 57:20). She hopes for agent-oriented languages like Bend, which combines programs with proofs. Proofs used to need a separate language such as Lean or TLA+ (Lauren Tan (poteto) 57:20; Lauren Tan (poteto) 58:11; Lauren Tan (poteto) 58:58).
12. Combining skill libraries and wrap-up [59:28]. Matt Pocock extends the conversation by five minutes (Matt Pocock 59:42). She recommends mining past transcripts for corrections and turning them into skills or lint rules, and says her Recall skill came from that practice (Lauren Tan (poteto) 1:00:35; Lauren Tan (poteto) 1:02:13; Lauren Tan (poteto) 1:03:01). She says everyone should own their own knives. In her view skills are malleable processes written in language, will get smaller, and can be mixed, for example Wayfinder with Potato Mode (Lauren Tan (poteto) 1:01:26; Lauren Tan (poteto) 1:03:52; Lauren Tan (poteto) 1:04:37). They agree to do it again (Lauren Tan (poteto) 1:05:16; Matt Pocock 1:05:17).

Follow-ups

• Matt Pocock and Lauren agreed to talk again. Lauren said “months” and Matt Pocock answered “Monday” (Lauren Tan (poteto) 1:05:16; Matt Pocock 1:05:18; Lauren Tan (poteto) 1:05:23; Matt Pocock 1:05:24).
• Matt Pocock is releasing a course on software factories. He films it next week and may change parts after this conversation (Matt Pocock 10:31; Matt Pocock 49:04).
• Matt Pocock planned to keep talking with Lauren after closing the stream (Matt Pocock 1:05:28).
• Lauren pointed viewers to her early skills project, poteto/noodle on GitHub (Lauren Tan (poteto) 3:32).

Open threads

• How to let agents merge their own code where many changes are one-way doors that cannot be reverted (medicine, law, finance). Lauren said it depends on how verifiable the domain is, admitted she has no answer, and said the industry still has to work it out (Matt Pocock 56:16; Lauren Tan (poteto) 57:20). (Lauren Tan (poteto) 56:28)
• Whether the setup is a “dark factory”. Lauren says it is, since agents merge overnight (Lauren Tan (poteto) 51:55). Matt Pocock contrasts it with Karpathy’s original vibe coding, where code barely exists, and suggests a dimmer switch (Matt Pocock 54:49; Matt Pocock 55:12). Lauren answered that the kitchen metaphor fits better, and the term was left unsettled (Lauren Tan (poteto) 55:23).
• Lauren lost her train of thought while explaining how a good environment helps human teammates, and did not return to the point (Lauren Tan (poteto) 46:43).
• Terminology: Lauren finds “software factory” apt but off-putting and prefers “Michelin kitchen” (Lauren Tan (poteto) 11:18). Matt Pocock still uses the factory term for his course (Matt Pocock 10:31).

Mentions

• Uncle Bob (Robert C. Martin) (person): Guest on Matt Pocock’s previous episode, about software quality and agents. (Matt Pocock 0:00)
• Cursor (now SpaceX AI) (organisation): Lauren’s employer since March, where she worked on the performance of the agents window. (Lauren Tan (poteto) 1:58, Lauren Tan (poteto) 4:20, Lauren Tan (poteto) 5:22)
• Meta / React team (organisation): Lauren’s previous job. She took a month off for burnout after leaving. (Lauren Tan (poteto) 1:58)
• Netflix (organisation): Lauren’s former employer, source of the management principle “context, not control”. (Lauren Tan (poteto) 40:49, Lauren Tan (poteto) 41:46)
• “How I shipped 2,500 PRs last month” talk (work): Lauren’s talk, posted on X about 10 days earlier, with about three million views. The Q&A is built around it. (Matt Pocock 1:01)
• pstack (transcribed P-SAC / PSAC / PStack) (tool): Lauren’s skill library. It includes verification, Recall, Potato Mode and full autopilot. (Lauren Tan (poteto) 3:32, Lauren Tan (poteto) 18:06, Lauren Tan (poteto) 52:49, Matt Pocock 59:44)
• poteto/noodle (work): Lauren’s early open-source skills project with a brain directory, which grew into pstack. (Lauren Tan (poteto) 3:32)
• GroqBot (also transcribed GraphBot; spelling uncertain) (tool): Personal agent with connectors that Lauren works on. It is her outer loop. (Lauren Tan (poteto) 5:22, Lauren Tan (poteto) 35:58, Lauren Tan (poteto) 42:09, Lauren Tan (poteto) 43:34)
• Cursor Projects (tool): New Cursor feature: coordinator agents in the cloud that delegate to sub-agents. It is Lauren’s inner loop. (Lauren Tan (poteto) 38:33, Lauren Tan (poteto) 42:09, Lauren Tan (poteto) 43:34)
• Dune (tool): Cursor’s internal framework for its Electron apps, like an internal Next.js, with strict lint rules. Not open source. (Lauren Tan (poteto) 29:38)
• Opus 5.5 (tool): Model Lauren says she loves, cited as an example of how capable frontier models have become. (Lauren Tan (poteto) 6:50)
• Andrej Karpathy / AutoResearch (person): Lauren cited AutoResearch for hill-climbing ideas. Matt Pocock used Karpathy’s original definition of vibe coding to describe a dark factory, which he contrasted with Lauren’s approach. (Lauren Tan (poteto) 18:06, Matt Pocock 54:49)
• Playwright and Chrome DevTools Protocol (tool): What Lauren’s verification CLI wraps. (Lauren Tan (poteto) 22:09)
• TypeScript / TSConf / Total TypeScript (work): Shared background. Lauren uses type narrowing as the model for constraints in a codebase. Matt Pocock noted TypeScript’s own 25,000-line God file. (Lauren Tan (poteto) 28:44, Matt Pocock 32:24)
• Bend (tool): Agent-oriented language that combines programming with proofs, Lauren’s hope for verifiable domains. (Lauren Tan (poteto) 57:20)
• Lean and TLA+ (tool): Separate languages for formal proofs, the older approach Lauren contrasts with Bend. (Lauren Tan (poteto) 58:11)
• Slack, Linear, X (tool): Outside sources of bug reports and context that the outer loop pulls in. (Lauren Tan (poteto) 35:08, Lauren Tan (poteto) 36:46, Lauren Tan (poteto) 42:09)
• Gordon Ramsay (person): Lauren’s example in the chain-of-restaurants analogy for scaling. (Lauren Tan (poteto) 34:17)
• Matt’s skills: Grill Me, Wayfinder (tool): Lauren suggested combining them with pstack’s execution skills. (Matt Pocock 8:48, Lauren Tan (poteto) 1:01:26, Lauren Tan (poteto) 1:03:52)

Quotes

I was the meat proxy between my agent and Chrome DevTools

Lauren Tan (poteto) 5:08. The bottleneck that started her verification work.

how do I make the easy thing the right thing?

Lauren Tan (poteto) 6:09. Her design principle for skills, since agents take shortcuts.

I actually feel like domain expertise is, is more important than ever.

Lauren Tan (poteto) 6:50. Her answer to Matt Pocock’s question about whether AI makes expertise less useful.

the single most important skill that should be in your toolkit is verification

Lauren Tan (poteto) 16:14. The central claim of the conversation.

I try to extract out the deterministic parts and turn that into code and just leave only the parts that actually require judgment to the agent.

Lauren Tan (poteto) 21:20. Her principle for building skills.

I almost feel like the new job of the engineer is really to, to spend time on the environment.

Lauren Tan (poteto) 26:14. Her view of the engineer’s role now.

How do I turn this into a lint rule? How do I make it so that the code base makes this impossible?

Lauren Tan (poteto) 31:16. How she responds to repeated agent failures.

you also, you know, for scale, you cannot be tasting every single dish that comes out of your kitchen

Lauren Tan (poteto) 49:31. Her answer on reviewing 2,500 PRs: sample them.

now I’m in a place where my agents are p- are merging their own code while I sleep.

Lauren Tan (poteto) 54:37. Where the trust ladder led.

The idea of sampling instead of blocking, I think is really important.

Matt Pocock 55:41. Matt Pocock’s summary of her approach to review.

The, the past chats I, I often say is like a, a treasure trove of context

Lauren Tan (poteto) 1:02:13. Her advice to mine your own transcripts for skills.

There’s nothing magical in them, right? They’re just words.

Matt Pocock 1:04:45. Matt Pocock agreeing with Lauren that skills are only language.

Statistics

Computed from the times of the transcript’s words. A turn runs until someone else takes the floor. A backchannel, a reply of at most 3 words and 1.5 s such as “yeah” or “mm-hmm” after which the speaker before it carries on, does not take the floor. The pause before speaking runs from the previous speaker’s last word to this speaker’s first.

Handovers. Each cell counts how often the row’s speaker was followed directly by the column’s.

The floor changed hands 62 times, 0.9 times a minute on average, and most often at 55:00, 6 times in that minute. The longest silence is 2.7 s, at 49:31.