


2,500 pull requests in a month. The README is the part to read
On 21 September 2026, SpaceXAI engineer Lauren Tan posted a 38-minute video with a caption saying she had shipped 2,500 pull requests the previous month, and the plugin behind it, pstack, opens by rejecting throughput as a goal.
By Katie Delaney / 2026-09-21 / 12 min read
Updated 30 October 2026: repository details and research checked.

What pstack is, and what Lauren Tan said
A fox counts the tracks before it trusts the trail. On 21 September 2026, Lauren Tan, who posts as poteto and works at SpaceXAI, posted a video on X under the words "here's how i shipped 2,500 PRs last month to production". She said it was meant for Cursor Compile in London, and that she missed it because she was livestreaming for Grok Bot Galaxy, so she released it free and told viewers to watch at 2x speed. The video runs 38 minutes.
The system behind it is pstack, and pstack is public. It is an MIT-licensed plugin in Cursor's plugins repository, copyright 2026 Lauren Tan, whose manifest listed version 0.15.36 when checked on 30 October 2026. Among Cursor plugins it is unusually opinionated. Its README opens by saying that "there's a growing sense that ai writes too much slop code", and that the author agrees.
The plugin's entry point is a skill called `/poteto-mode`. Around it sit twenty-three playbooks for different kinds of task, plus engineering principles and specialist agents, as listed in the repository on 30 October 2026. Other skills include `/swarm`, which the README describes as "N parallel workers across different slices or races, then one aggregated report", as well as `/tdd` and `/unslop`. pstack works with Cursor's own `/loop` command, so an agent can run for many hours. `/loop` belongs to Cursor, not to pstack, and I could not find a `/goal` skill in the folder.
The plugin is also moving fast. Flavio Copes's deep dive from 21 August counted twenty-one principles at the time, and notes that the same skill files load in Claude Code and Codex.
One design choice matters for anyone building agentic engineering workflows of their own. pstack routes work by kind. The README says that "code delegates (feature, refactoring, bug fix, perf, hillclimb) go to grok", while the hardest changes, prose and judgement default to Opus 5.5, and the default panel is Opus 5.5 and Grok, according to the README on 30 October 2026. The setup command lets users change these choices. For context on model comparisons, see our Grok 4.7 scorecard piece.
The number has already moved
The headline figure is worth a second look before it is repeated. Cursor's Compile London page lists her afternoon slot as "I Shipped 2,000 PRs Last Month". Her post caption says 2,500. The figures differ, and the sources reviewed do not establish why. Both are self-reported and we found no independent audit of either. A count of pull requests tells you how much was opened and merged. It does not tell you how much stayed merged.
The README's rule: go deep first
Read the README before the headline and the two pull in different directions. "Throughput without quality is not a goal i aspire to", it says. "If you want to go fast, go deep first." It also says, "personally, i don't believe in planning. the best spec is code", which is an unusual thing for a plugin author to put in the front door. pstack is not a factory for pull requests. It is a discipline for making each one worth merging.
The discipline is verification. The prove-it-works principle tells the agent to verify every task output "by checking the real thing directly" and not to "infer from proxies, self-reports, or 'it compiles.'" A companion rule for delegation is short enough to remember: trust artefacts, not self-reports. An agent that says it is done has told you nothing. A test that ran and a value that was read tell you something.
here's how i shipped 2,500 PRs last month to production
Her view of where this leads is stated plainly in another post from 19 September, about the Bend programming language: "when you have a codebase that can formally verify itself, you can ship at an incredible pace. code review is solved". Read together, the position is consistent. Speed comes from automated proof, so human line-by-line review becomes the exception. Whether a given codebase can carry that weight is the open question.

That is the claim worth testing, and it is testable. A team using pstack, or any agentic engineering setup, can measure the rate at which agent pull requests are reverted, the share merged without any review, and the time humans still spend reading diffs. Tan has not published those figures alongside the 2,500, at least not in anything I could read without logging in to X.
| Claim | Status | What I found |
|---|---|---|
| Shipped 2,500 pull requests last month | Self-reported | Her own post; no independent audit |
| Compile London listed the talk as 2,000 PRs | Confirmed | Cursor's schedule carries that title |
| pstack includes /loop and /goal | Partly wrong | /loop is a Cursor command; no /goal skill in the folder |
| A Chief of Staff agent runs the bots | Not in pstack | Absent from the README; appears in reposts about Grok Bot practice |
| 85% of SpaceXAI engineers run 10+ bots | Secondhand | Repost only; no first-party source found |
| Is on the React compiler core team at xAI | Confirmed | Her GitHub bio says so |
- Shipped 2,500 pull requests last monthSelf-reportedHer own post; no independent audit
- Compile London listed the talk as 2,000 PRsConfirmedCursor's schedule carries that title
- pstack includes /loop and /goalPartly wrong/loop is a Cursor command; no /goal skill in the folder
- A Chief of Staff agent runs the botsNot in pstackAbsent from the README; appears in reposts about Grok Bot practice
- 85% of SpaceXAI engineers run 10+ botsSecondhandRepost only; no first-party source found
- Is on the React compiler core team at xAIConfirmedHer GitHub bio says so
The last row comes from her GitHub profile. Two of the others deserve a note. The Chief of Staff idea is real in the sense that reposts describe it as a Grok Bot team pattern, but it is not part of pstack as published, so anyone installing the plugin expecting it will not find it. And the 85% figure has no source I can trace to Tan or to SpaceXAI. It may be true. It has not been shown.

What independent data says about AI coding agents at volume
A single engineer's month is an anecdote, and an impressive one. To judge the pattern, look at larger samples. The evidence is mixed, which is the honest summary. It does not say agents are worse than people. It does say volume by itself is a poor scoreboard.
Start with the sceptical end. METR ran a randomised trial with sixteen experienced open-source developers across 246 issues and found that with early-2025 AI tools they took 19% longer than without. Those were assistants, not today's agents, and the sample was small, so read it as a caution about measuring speed, not a verdict on agents.
METR later cautioned that its follow-up could not reliably estimate the current effect, because participation and task-selection biases had grown. The early-2025 finding is not a measurement of today’s tools.
Now the telemetry end. Faros AI's AI Engineering Report 2026 covers two years of data from 22,000 developers. It reports that the incidents-to-pull-request ratio is up 242.7% as teams move from low to high AI adoption, and that pull requests merged without any review are up 31.3%. Faros sells engineering analytics, and its numbers are correlations, so treat them as a warning light, not a verdict.
A separate 2026 longitudinal study examined 669 developers across 103 repositories and a control group. Review became faster and less detailed in both groups, so its authors did not attribute that change to agent adoption. It did not test pstack.
rise in the incidents-to-pull-request ratio between low and high AI adoption, across 22,000 developers, per Faros AI's 2026 report
Which agent matters more than whether it is an agent
The best-designed evidence in the middle is an academic study of agent pull requests, posted to arXiv on 12 September 2026. Researchers followed 37,623 pull requests across 2,807 GitHub repositories, written by five commercial coding agents between December 2024 and July 2025, and compared their post-merge behaviour with a human baseline. That window is before today's models, so treat it as a method, not a scoreboard. The paper found that quality differences are vendor-specific, not uniform. Codex pull requests were reverted about half as often as human ones, at 6.1% against 11.5%. Devin pull requests were reverted more often, at 14.5%.
The same paper found that agent code, pooled across vendors, was less likely than human code to contain a security smell. So the picture has two sides. Agents can be safer on some measures and riskier on others, and the spread between agents is wider than the gap between agents and humans. That matters for anyone reading pstack's model routing, which defaults code to Grok and harder judgement to Opus 5.5, with user-configurable choices, as a claim that the choice of model is part of the quality system.
Put the three findings beside Tan's story and a sensible reading emerges. Her setup is built around verification, which is what the risk data says to prioritise. The risk lies in copying the volume and skipping the proof. A team that adopts ai coding agents for throughput, keeps line-by-line review, and does not add automated checks gets the cost of both approaches.
What a team can copy from pstack without the headline
Strip out the number and pstack offers a practical shape for agentic engineering that works for any team, including the marketing and product groups folkfox advises. The plugin is a set of habits written down as files. Habits travel further than a plugin does, and each of the five below can be adopted this week.
Write down what proof looks like for each task: the test that ran, the value that was read, the diff that was inspected.
Never accept an agent's own statement that work is finished. Ask for the artefact and check it directly.
Track merged work that is still merged after a week, plus reverts, not pull requests opened.
Send routine code to a cheaper model and keep judgement and prose for a stronger one, as the README does.
Measure how much lands with no review at all. If it is rising, the verification is not keeping pace.
The second habit is the one teams skip. It feels slower, and for a week it is. But it is the difference between an agent that reports success and a system that has succeeded. A useful test is to ask what a colleague could check in ten minutes without asking the agent anything.
None of this needs a specific tool. It needs a definition of proof, a scoreboard that rewards it, and the discipline to keep the humans on the judgement calls. If you want help designing that for a team using Cursor plugins or any other agent stack, folkfox's AI consultancy work starts with exactly those three things, and the content marketing services team applies the same checks to AI-drafted copy before it ships.
For AI code review specifically, pstack's `/interrogate` skill, which the Flavio Copes guide files under multi-model review, is a good template. It takes a diff and puts it in front of several different models, and the logic is that one model's blind spot is often another's obvious catch. It is not a replacement for a human on a risky change, but it is a cheap second reader, and a second reader is the theme of this whole story.

The bottom line on pstack and 2,500 pull requests
Tan has done something useful in public. She wrote a workflow down, released it under a permissive licence, and explained it in a free video. That deserves credit, and the README shows real thought about quality. What is missing is the audit. The 2,500 figure is her own, the Chief of Staff and 85% claims are secondhand, and the independent data says volume without verification is where the risk sits.
Grok Bot, the product behind the livestreams, is described by SpaceXAI as "your team of always-on agents" that "keep working 24/7", launched on 11 August 2026. Always-on agents raise the stakes for exactly the habits above. An agent that never stops can also never stop being wrong, unless something checks it. Earlier folkfox pieces on the Grok Bot Galaxy workflow test and on an agent refactoring a million-line codebase make the same point from different angles.
What to watch next is simple. Does anyone reproduce the 2,500 with an audit trail, does the pstack repository show how it defines merged and verified, and do independent boards or studies pick up its methods? The folkfox news desk will revisit if they do. Ask for the number that survived a week, not the number that landed on a Monday.
Frequently asked questions
What is pstack?
pstack is an open-source plugin for Cursor by Lauren Tan of SpaceXAI. It bundles a main skill called poteto-mode with twenty-three playbooks, engineering principles and specialist agents, according to its README on 30 October 2026. Its stated aim is rigorous agent workflows you can parallelise with confidence, built on verifying work directly.
Is pstack free to use?
Yes. pstack is released under the MIT licence, with copyright held by Lauren Tan, and lives in Cursor's public plugins repository on GitHub. The licence allows use, copying and modification. The models it routes work to may carry their own costs.
Did Lauren Tan really ship 2,500 pull requests in a month?
That is her own figure, from the caption of a video post she shared on 21 September 2026. Cursor's Compile London schedule listed the same talk as 2,000 pull requests. Nobody outside her team has published an audit of either number, so treat both as self-reported.
Does pstack include a Chief of Staff agent?
Not in the README or the guide pages I read. Chief of Staff appears in reposts describing how people run Grok Bot teams, where one agent coordinates others. It is a real pattern in that setting, but I found it in no pstack skill. The plugin does include a swarm skill for parallel workers.
What is Grok Bot?
SpaceXAI launched Grok Bot on 11 August 2026 and describes it as a team of always-on agents with their own computer that work inside tools and apps and keep working around the clock. Grok Bot Galaxy was a three-day event of demos and build sessions in September.
Does AI agent volume lower code quality?
The evidence is mixed. Faros AI's 2026 telemetry links high adoption to more incidents per pull request and more unreviewed merges. An academic study of 37,623 agent pull requests found quality varied by agent, with Codex reverted less often than humans and Devin more often.
Grok 4.7's price is real. The scorecard needs a second reader
Today's other SpaceXAI story, checked against public leaderboards.
Read the pieceAI consultancyGrok Bot Galaxy: 5 essential AI workflow tests
The workflow controls to watch when agents build a company live.
Read the pieceAI consultancyHermes Agent refactored a million-line codebase. The audit matters more
Why the audit trail counts for more than the size of the change.
Read the pieceAI consultancyThe cheap coding agent that is almost frontier
How to test a near-frontier model on your own work before you switch.
Read the pieceRunning agents on real work? Define proof before you count output.
folkfox helps growth and product teams set up agent workflows with verification built in, so the number you report is the number that survives.
Want folkfox in your Google results and AI answers? Set folkfox as a preferred source.

