Poteto's Trust Ladder: Let the Agent Verify Itself Before You Let Go

🌏 中文版

Start the English video exercise (3 clips, 2 min 8 sec; partial bilingual captions)

If you already use a coding agent but still have to watch every conversation, this post asks one thing: which parts of Lauren Tan's approach can you try tomorrow, and which depend on conditions only she has?

What this is

Lauren Tan (poteto online) was on Meta's React team and now works on Cursor's agents window and Grok Bot, both under SpaceXAI. On October 2, 2026, Matt Pocock hosted a 65-minute livestream conversation with her about producing a lot of work with agents without losing quality. Her skills are public under the MIT license as pstack, entered through /poteto-mode.

Original interview on YouTube. Use the practice player below to listen here.

The number first: 2,500 is self-reported

The PR count differs by source:

• The stream title says "1,000's of PRs".
• Her September 21 X post says 2,500.
• Cursor's Compile London schedule lists the September 16 event’s talk as "I Shipped 2,000 PRs Last Month".
• An analysis of the Compile talk's auto-captions says the spoken figure is about 2,000.

All are her own numbers for a month she chose, with no outside audit and no data on how many merged PRs were later reverted. At about 45:54 in the conversation, she says a good share are small "gardening" fixes. So the figure tells you how finely she slices work, not how much value shipped.

Who it is for, and prerequisites

It fits people who use coding agents daily and find themselves the bottleneck. You do not need Cursor or pstack; the transferable part is the thinking. The prerequisite is that something in your project can judge right from wrong by code: a runnable app, tests, types, lint.

Content map

1. The trust ladder. Connecting her kitchen metaphors across different parts of the conversation gives this reading: one person watching one agent is a home cook; adding agents without a system is a family crowding into the kitchen, which only makes things worse; then comes the head chef who designs how the kitchen runs; finally the owner who moves between branches. She dislikes the associations of "software factory" that downplay quality and craft, and separately reminds us that the engineer’s reputation remains attached to the work.

2. A verification skill is the first rung. At Cursor she started as the "meat proxy" between the agent and Chrome DevTools: reading flame graphs and heap snapshots herself, then relaying them. Her first skill after joining Cursor let the agent launch the app, use it, and capture traces. Her definition of a loop follows: it is a loop only if the agent can check its own work.

3. Deterministic work as code, judgment for the agent. The verification skill contains a CLI wrapping Playwright and the Chrome DevTools Protocol. She says herself it is not impressive; the point is that every agent shares one toolset instead of rewriting a verification script each time, and behaves consistently. Migrations follow the same rule: prefer codemods over an agent reasoning about each file.

4. Environment over nagging. Each time an agent errs, she first asks how to make the mistake impossible to write; the answer is usually a lint rule, a type, or a directory convention. She describes Dune as a non-open-source internal framework: each feature has its own directory, a registry and codebase scanning discover features, and restrictive lint rules narrow the available patterns. The transferable part is the design idea.

5. Outer and inner loops. The inner loop is the agent writing code from your intent; the outer loop is the new information constantly arriving in Slack, Linear and X. Her personal agent subscribes to those sources and hands work to a cloud coordinator that splits tasks among sub-agents. Giving related bugs to one coordinator makes it easier to spot the shared root cause.

6. Sample after merge. The flow: an agent makes a PR, several verifiers exercise the running app, issues get fixed, and it merges automatically; she reads the commit history the next morning. When several agents take the same shortcut, she fixes the skill, lint, or types, not the one agent.

One concrete example: which layer should a correction live in?

The official pstack guide’s /correct section gives four places to fix recurring mistakes, preferring the strongest enforceable option:

1. Architecture or data structures that make the mistake impossible.
2. Types, lint or CI checks that block it and explain the fix.
3. Tests that catch it.
4. Documentation or agent rules when the earlier options do not work; skipping text alone does not fail a check.

This is the current public tool’s correction strategy, rather than a ranking recited in this interview. /correct also requires proving a new check fails on a real historical mistake, so adding a plausible rule is not enough.

flowchart TD
    A[Agent repeats a mistake] --> B{Can architecture or types<br/>make it unwritable?}
    B -->|Yes| C[Change architecture and types]
    B -->|No| D{Can it be a lint rule or CI check?}
    D -->|Yes| E[Add a lint rule]
    D -->|No| F[Find a test that catches it<br/>then consider docs or rules]

For recurring errors, look for a fix in code or checks before falling back to written reminders. The official guide's starter example is similarly plain, a goal plus a way to check it:

/poteto-mode the export writes duplicate rows when a retry lands mid-run. repro first, then fix and verify.

Source: the pstack guide.

Limits

• Scope: Matt asks about irreversible changes and domains such as medicine, law and finance. She says programmatically hard-to-verify domains make it harder to reach this level of autonomy, and acknowledges she has no complete answer. Her metaphor of verification turning one-way doors into two-way doors does not mean verification can undo data loss. This article recommends retaining human merge judgment when verification is insufficient or consequences cannot be reversed.
• Cost: Full autopilot spawns several verifiers per PR, and she admits it burns tokens.
• Heavy prerequisites: Her codebases were deliberately narrowed to one way of doing things; an early Grok Bot had eight files of 10,000+ lines each and was split under pressure. Applying this to an existing project still requires building corresponding tools and constraints.
• Verification must cover the goal: Passing existing functional checks does not cover every quality or performance target. She also uses verification tools for performance improvement; those goals need appropriate traces, profiling or benchmarks.
• Sourcing: The main claims have been checked against automatic English captions; the practice excerpt has not been checked by listening to the audio. Dune is the internal framework she describes, and the PR numbers and workflow are her own account. For her exact words, go back to the original stream.

How to use it: three things for tonight

1. Pick the app you change most and write a command that lets the agent launch it, run one main flow, and leave evidence (a screenshot or trace); have it run after every change.
2. The second time an agent makes the same mistake, stop and ask whether it can become a lint rule, a type, or a directory constraint. Editing the prompt is the last resort.
3. Go through your past chat logs, pick the three things you keep correcting, and write each into a skill or rule. She also advises building your own set rather than copying hers wholesale; Matt's skills repo is a reasonable template.

For why harnesses matter, pair this with the site's Phil Schmid guide and the meta-harness layering post.

Listening and speaking practice

Browse more English video exercises. Use the three clip buttons above to switch within the same player. Play the whole clip, then click a quoted line to repeat it; choose 0.75× if needed. The selected lines have Chinese translations; the player’s CC provides the remaining source captions.

Each clip includes a listening goal, a language note and a speaking task. The examples are written for this site and are not the speaker’s words. Clip boundaries follow automatic-caption cues and have not been checked by listening; the first and last words can continue outside the selected range.

Update history

• 2026-10-20: Rebuilt 548 timed semantic units before translation; source-word coverage retained. Selected passages now share sentence boundaries with the full transcript, which follows playback without requiring prior selection.

• 2026-10-20: Imported all 1707 supplied subtitle segments with Traditional Chinese machine translations and sentence replay. Original clip exercises retained; translations have not been reviewed sentence by sentence.

• 2026-10-19: Added a listening goal, language note and authored speaking example for each clip; added two brief caption excerpts and consolidated duplicate players. These remain selected quotations, not complete bilingual transcripts.

• 2026-10-19: Combined three original-audio clips into one switchable player and labelled caption coverage. Complete bilingual captions for all three clips remain unfinished.

• 2026-10-19: Added a guided clip exercise with revealable comprehension answers, contextual phrases and an original speaking task; added the English video practice tag.

• 2026-10-10: Labeled the exercises as shadowing with source-player captions; distinguished short quotes and paraphrased notes from complete bilingual transcripts, and added timed source/transcript instructions.

• 2026-10-10: Added three continuous 31–51-second topic clips with synchronized English captions, comprehension prompts and a listening-to-shadowing progression; kept the short quote as a locator.

• 2026-10-10: Checked the main claims against original interview captions; clarified Dune, verification coverage and irreversible changes; corrected the Compile event date and replaced the correction ranking with the official pstack /correct guide.

• 2026-10-10: Replaced isolated phrases with a coherent caption excerpt, corresponding audio player, English captions, playback range, and practice steps.

• 2026-10-10: Retrieved automatic English captions and added short excerpts, Chinese comparisons, and original speaking exercises. Caption text checked; audio verification pending.

References

• Matt Pocock × Lauren Tan livestream (YouTube, 2026-10-02)
• Lauren Tan's X post: how i shipped 2,500 PRs last month
• pstack (cursor/plugins, MIT)
• pstack guide
• pstack guide: Make it yours (/correct enforcement levels)
• Cursor Compile London schedule
• Agile 3 Uncles: the secret of 2,500 PRs a month (in Chinese, secondhand)
• PJFP: Poteto on pstack (chapter summary, secondhand)
• Redreamality: Trust First, Then Parallel (Compile talk caption analysis, secondhand)
• Matt Pocock's skills repo