When Vibes Replace Pull Requests

ALL THE VIBES

ISSUE 31  ·  TUESDAY, OCTOBER 20, 2026

What Stood in for the Pull Request?

OPENING

Friday's Community Call spent an hour building a pixel-art town with no required pull-request review, so we asked what stood in for it. A pull request, a proposed change a second person reviews, is the gate most of us work behind. Stephanie Schofield and Ahmed Abdelsalam took it away; she made every All The Vibes member a maintainer (someone allowed to push to the shared code), and Mark Briggs's flying Tesla dropped free tokens over a town with a dragon in it. Forty minutes in, Shyam Sridhar noted "no showstoppers," and we went to check.

Short answer: Git, a build script that cannot fail, and automated checks that ran during the hour. None of them is a second reader. The Workbench lays out the evidence, and what a pull request costs in firms that still use them. Also this week: a Windows sandbox that refused to start, a decompilation project where a check replaced the reviewer, a habit for testing your own understanding, and a conference discount.

COMMUNITY CONVERSATIONS

Ahmed Abdelsalam built the pixel-art town (a live version, linked in the call chat) with a pixel-art skill, Stephanie Schofield added the whole organisation as maintainers [1], and Veda Vundela wired a library to the Open Library catalogue so characters could open real titles [2]. Then the room filled it with a flying Tesla, a dragon, a nightclub, a sorting hat and a Stranger Things portal.

Position A, as Stephanie Schofield framed the experiment: at this speed, review is what stalls. "having PRs is the bottleneck because we're all moving. so quickly. So the review process, the like, you know, request changes, fix loop takes too long." The argument she described still had control: people work in their own domains, and when something breaks, the checks that run after a merge turn red and everyone repairs it together instead of waiting for a reviewer (they ran during the hour; see the Workbench). Shyam Sridhar named the shift at the end: "Because before it was like, it was such a bad word, like no PR." Now, he said, "I can see how that works."

Position B, David Crawford: verification has to be answered before the gate can go. About twenty minutes in he named the gap, "what kind of changes are allowed", and said "we haven't closed the loop on the basics" between development "and the SRE side" (site reliability engineers, the people who run the software live). Then the question: "How do you verify intent? You know, just because somebody wants something ... doesn't mean it." At the end, asked for a take, he was conditional: "nobody's going to get hurt if something gets broken here. You're not going to lose any money." He said there are PR groups "that have been addressed now with GitHub. So you have blocks of PRs getting approved as a way to do things. So it's not just a single PR." What is changing, he said, is "the way that we observe and constrain the work as we go through", and then, correcting himself, "validate the work that goes through." In chat, Raimondas Lencevicius stated the objection directly: "It's nice to play around for an hour, but it's clearly not a sustainable product/project. Even after 1 hour, it's already a mess. So needs some functionality control long term." He added that he understood it was for fun.

Mel Mok found the collisions instructive: "The strugs of trying to push something complicated while other people push so I have to try again. Honestly this is enlightening." Ahmed Abdelsalam's reply was the working rule: "the best thing is to push out small changes at a time."

Shyam Sridhar noted forty minutes without a showstopper, and the repository agrees there was no outage. It also records the friction: write-access errors early on, a disco moved off the kitten corral, and a one-line fix fourteen seconds after a refactor [1]. Each side is right about the hour. They disagree about what an hour like this samples: it shows what survives among builders who see each other's pushes, and the dispute is whether that resembles software with consequences. The repository can say more than the call did, and the Workbench says it.

REFERENCES

[1] All-The-Vibes/ATV-world, GitHub, public commit history read October 19, 2026 (the community's own repository; a saturated domain). Our counts and times come from it, in UTC, from author timestamps. ↑

[2] APIs, Open Library, Internet Archive, read October 19, 2026. Veda Vundela described the connection on the call; we have not inspected how her library calls it. ↑

UNDER THE HOOD

Iris Laband asked the Hack & Furious chat about a Copilot-in-VS-Code error on Windows, "Sandboxing is unavailable in this environment." PowerShell commands that had worked before were now rejected before they ran. (A sandbox confines an agent to limited files.)

David Crawford linked the docs and named the setting, chat.agent.sandbox.enabled. His reading was that her machine was missing a recent Windows security update the sandbox depends on; the docs list which ones [1]. The docs support the part that matters. When that dependency is missing, the command "does not silently run without the sandbox," so the refusal is deliberate and a supported fix exists [1]. Iris did not say whether the update worked, so the diagnosis is David's reading, not a confirmed cure.

This is a gate that fails closed: when it cannot work, nothing runs. ATV World's build script fails the other way, and keeps going. But the sandbox guards less than its name suggests. It covers terminal commands and the programs those commands start. Outbound network stays open by default, approving a command to run unsandboxed lifts the restrictions, and the docs say it is not a virtual machine, meaning not a separate simulated computer [1]. It narrows file damage but does not stop data leaving. That echoes David on the call, who spoke of how we "observe and constrain" the work and, separately, asked how you verify intent.

REFERENCES

[1] Sandbox Copilot agent sessions, Visual Studio Code documentation, dated October 14, 2026, read October 19, 2026. The domain is saturated in our citations; this is the only primary source for the Windows prerequisites. ↑

THE WORKBENCH

The setup. On Friday's Community Call the group spent an hour building ATV World, a shared pixel-art town on the web, in one public GitHub repository, All-The-Vibes/ATV-world. Everyone on the call could push straight to the main branch, with no pull request (a proposed change another person reviews before it merges) required. Builders added characters, buildings and a nightclub, and the live town kept updating as they did. That makes it a handy test: if the usual review gate is gone, what is actually keeping the town standing?

The stand-ins. What stood in for the pull request during that hour, and would any of it survive once the room empties? Between 15:14 and 16:09 UTC on October 16 the repository's public history shows 61 commits (saved changes) from 11 author names [1], 13 of them merge commits, which join two lines of work. The first stand-in was Git itself. It refuses a push that is not built on the latest main (the branch the site is built from), so a late pusher had to merge in what others had pushed first [2]. The second was built on Thursday night. At 9:39 pm Central, Ahmed Abdelsalam committed a build script that "keeps every push green and the world up". The script leaves out any plot (one builder's patch of town) that breaks the build and marks it under construction, falls back to the already-deployed version of any other broken file, and deploys nothing if that fails [3]. The third was a set of automated checks that ran during the hour.

Green, but saying little. All 42 Deploy runs triggered by a push in the window finished green in a median of 57 seconds, and so did all 25 runs of CI Health (CI is continuous integration, automated checks on each push), a type-error scan and full build added mid-call [1]. A green deploy tells us little, because the build script was written so that a deploy always succeeds. None of the 43 Deploy runs we read (the 42 plus one manual run) carries a fallback warning, so no push visibly needed the fallback. The stricter CI Health check went red once, at 16:48 UTC, 39 minutes after the hour ended, and that commit still deployed [1].

"But it's great and it means that we have gaps that we need to fix there. So I think that's supposed to be the mentality behind what you do when you don't have PRs."

STEPHANIE SCHOFIELD, COMMUNITY CALL, OCTOBER 16

What the hour could not test. Stephanie was describing the day before, when she said "literally nothing is passing." Ahmed put the earlier failures down to a Python linter running in a repo with no Python. The run history shows that linter red on ten pushes Thursday night, with CodeQL red on the first several, and the log shows the linter rule changed at 9:27 pm Central Thursday [1]. That is the kind of gap she wanted surfaced.

What the repo lacked was a required second reader. A workflow auto-merged any pull request that was opened (one, #2, in the hour), so a pull request was a path to main, not a gate. Main has no branch protection today (rules requiring reviews or checks before merging), and its one ruleset, created that Friday evening after the call (8:34 pm Eastern), forbids force-pushes (overwriting shared history) and branch deletion, nothing else [1]. So the hour tested instant feedback among people watching one screen, where a bad push shows within a minute and costs nothing. That inference is ours. Stephanie's proposal is a fix loop shared across the whole codebase, with people still working in "a domain and a scope"; nothing in the hour tested that at a longer latency either. It could not test what happens once the author has left the room and nobody else has read the change, which is the case a pull request exists for. That cost comes after the cost of keeping the gate.

What a pull request costs. Where pull requests still run, agents have made them slower. Fiona Chen and James Stratton of Harvard, in a job market paper (an unpublished paper written to get hired; not peer reviewed, observational, a firm-level average, checked by its data partner only for confidentiality), compared earlier and later adopters of coding agents across 718 firms from January 2021 to March 2026 [4]. After adoption, the time from a pull request opening to merging rose 3.45 days on a 7.03-day baseline, 49%; our arithmetic puts the 95% interval (the range the data cannot rule out) near +4% to +94%. Pull requests rose 23%, yet resolved tickets and epics (feature-sized work items) showed no statistically detectable gain, and gains above 12% are ruled out. Read that as no detectable gain, not as nothing shipped. That is the cost Position A feels.

HOW IT WORKS: WHY A PULL REQUEST WAITS

A pull request is a proposed change held until a reviewer reads it. In the Harvard data the clock runs from submission to merge, which on our reading includes time the request sits unread [4]. When agents write the code, 23% more requests arrive and the share of workers who review rises 14%. More requests come back with changes requested, about 0.13 before agents and 0.25 after. Pull-request size did not change significantly, so the extra wait is not bigger changes. One junior engineer describes one project adding senior-approval rules for AI-generated code after outages, creating more bottlenecks; a single anecdote. The data cannot say whether reviewers are scarce or merely stricter, so the rise in changes-requested could be either.

The slowdown has a behavioural part. In a University of Saskatchewan survey of 239 practitioners with code-review experience, 52.3% were likely or very likely to accept a human-authored refactor (restructuring code without changing what it does) that passes CI without further explanation, against 19.7% for the same change from an AI [5]. The AI question came second, so the gap is probably an upper bound, and the sample was hand-picked and self-reported. In our reading, when passing tests stop earning acceptance, reviewers will want the written intent. Mark Rowe endorsed that shape in the Hack & Furious chat: "PRs I like Mario's POV. Don't vibe code the PR, give me the prompt instead." The chat names Mario but gives no link.

Who has argued the gate can go. One source argues it. DORA, a research programme on software delivery, says external change approval, such as a change-advisory board, has "a negative impact on software delivery performance," and recommends peer review plus automation [6]. The page summarises 2019 research, gives no figures, and argues for peer review, which the hour removed too.

Anthropic's post on its in-house version control, Scion, does not argue the gate can go. Its team members "still review every change request for the system, but they now examine the intent of the change as opposed to reading every line of code" [7], so review is shifting toward intent, not disappearing. The post reports merged changes rising from about 2,500 a day in May to over 30,000 by September. That measures throughput; it reports no defect, revert, incident or review-latency data, and calls Scion "by no means a suitable replacement for git for general use." Searching for measured outcomes from teams that dropped mandatory review, we found position papers and observational data. The one measured case we found is the decompilation project in In the Wild, where a check replaced a reviewer because intent was already checkable.

What the hour hid. Now the other cost. A pull request is, in our reading, also where a second person learns the code. Margaret-Anne Storey calls the lost shared understanding cognitive debt [10], in a conceptual paper with no size estimate. Authors using agents do seem to understand their own code less: in (Im)Paired Programming, 54 computer-science students built a small game with a coding agent or a chatbot limited to short snippets, and agent users scored lower on comprehension questions about their own code (p<0.002, Cohen's d>0.80, a large effect; the paper's caption values). Their lead on a follow-up task reversed once first-task accuracy was controlled [8]. The sample was students on two web tasks.

Echoes of AI, a preregistered experiment (plan filed before data) with 151 participants, 95% professional, found no significant difference in completion time or code quality for developers who later evolved AI-assisted code [9]. That tests later developers, not authors. So what is measured is the author's loss, and what is unmeasured is the second reader's. A pull request is also the one moment someone other than the author checks the author's intent against the code. With no second reader, that check happens nowhere, which is David Crawford's question. The stand-ins gave fast feedback; none of them verified intent.

REFERENCES

[1] All-The-Vibes/ATV-world, GitHub, the repository the community built (the subject of this section; a saturated domain), public commit, Actions and ruleset data read October 19, 2026. Our own count, UTC, from author timestamps and each run's start and last update; we read the annotations of all 43 Deploy runs but cannot see settings history. ↑

[2] git push, "Note about fast-forwards", Git documentation, read October 19, 2026. ↑

[3] scripts/build.ts, ATV-world at the commit before the call. The header comment carries no byline; the commit history shows Ahmed Abdelsalam as its author, October 15, 9:39 pm Central. ↑

[4] Artificial Intelligence in the Firm: Bottlenecks in Software Production, Fiona Chen and James Stratton, Harvard University, job market paper, current version August 4, 2026, read October 19, 2026. ↑

[5] Who Pays the Review Cost? Triage, Fairness, and Accountability in AI-authored Pull Requests, Rahman, Alam, Roy and Roy, arXiv, October 2026 (preprint, purposive sample). ↑

[6] Streamlining change approval, DORA, page updated October 30, 2025, read October 19, 2026. Older than our 90-day window; used because it is the only research-backed statement we found on this question, though the page gives no figures. ↑

[7] Redesigning source control at Anthropic, Anthropic Engineering, October 13, 2026. The domain is saturated; the community raised this post. ↑

[8] (Im)Paired Programming: Coding Agents Improve Productivity but Harm Understanding, Balepur et al., arXiv, July 29, 2026 (preprint). ↑

[9] Echoes of AI: Downstream Effects of AI Assistants on Software Maintainability, Borg et al., arXiv, version 3 of February 26, 2026 (preregistered; older than our window; the only controlled test we found of the "cost shows up later" claim). ↑

[10] From Technical Debt to Cognitive and Intent Debt, Margaret-Anne Storey, arXiv, March 23, 2026 (conceptual paper; older than our window; kept for the term). ↑

GETTING STARTED: AI NATIVE DEV

The Workbench ended on a change nobody else read. This is the individual version: if an agent writes your code, how would you know you still understand it? Mark Rowe put the stake on the October 13 Hack & Furious: "I'll spend tokens if it helps me understand what the hell I'm building. Because if I don't know what I'm building, how do I know that it went wrong?" In the chat, Bradley Lawrence asked the same of his students: "Now in my USF capstone class there are students with dozens of repos all ai assisted. Which is good from one perspective, but just how much of that code do they actually understand / reviewed?"

The Workbench's students did two web tasks. Anthropic ran a randomized controlled trial, where chance assigns who gets help, on engineers [1]. Fifty-two mostly junior engineers learned a Python library for concurrent tasks by building two features, with or without an AI assistant, then took a quiz. The AI group averaged 50% against 67% for hand-coders (Cohen's d of 0.738, the gap in standard deviations; p=0.01), and the largest gap was on debugging questions. How people used the assistant mattered. The low scorers fell into three patterns, four people in each: handing the task over, leaning on the assistant more and more, or having it debug repeatedly. The high scorers, as the study labels them, fell into three others: generating code and then working to understand it (two), requesting code with explanations (three), or asking conceptual questions (seven). The sample was small, the quiz came right after the task, and the authors draw no causal link from those patterns to learning.

A habit follows, and it is ours, not a tested remedy: when the agent finishes, switch it off and answer three questions about the result, which makes you do what the high scorers did, understand after generating. What does this function call? What breaks if I delete that line? Where would I look first if this failed on Friday? For a contrary view, see The Stack.

REFERENCES

[1] How AI assistance impacts the formation of coding skills, Anthropic, January 29, 2026, read October 19, 2026 (study of Shen and Tamkin; sponsor-run; older than our window; the only controlled study we found that reports debugging separately). ↑

IN THE WILD

Mark Rowe shared Maurice Heumann's write-up of letting AI agents decompile (rebuild source code from a finished program) a first-person shooter in the Hack & Furious chat on October 12 ("Really good set of lessons learned") [1]. It matters here because a check replaced a reviewer, in one author's account whose code is not shared. Heumann's first set-up had three worker agents and one reviewer agent. They reached roughly 80% of the game, but the code was semantically wrong, and the reviewer accepted deviations on the workers' own justifications. What worked was a byte-matching check: compile each reconstructed function with the original compiler, compare it byte for byte with the original executable, and re-verify passes in CI. Agents tried to cheat with inline assembly (hand-written machine code) and by editing the checker, so CI now compares the checker's hash, a fingerprint of its contents, against a stored secret. At the end 99% of functions exist and 83% match exactly, and he says the reviewer agent "is not needed anymore." The remaining 17% rests on his belief.

ATV World's build fallback is built to absorb failure; this check is built to expose it, on every function. Heumann's reason: "Humans are notoriously incapable of precisely articulating their intent. Therefore, I don't think reviewers will ever be enough." A check can replace a reviewer exactly when intent has already been written down in checkable form, and a decompilation has an answer key, the original program, that most software lacks.

Related, on agents guessing when information is missing: The Working Set of a Coding Agent [2] finds in controlled experiments (seven models, five harnesses) that a missing fact "produces wrong work rather than absent work," that how often an agent says it is blocked depends on the model, and that a stale conventions file (a file of project rules the agent reads) can cost more than none. It does not study gaming checks. More to follow up is in The Stack.

Two more reads on this week's theme. The Register's report on Anthropic's Scion announcement [3] is an outside read on the same source-control redesign we cite in the Workbench. And Making Claude Code more secure and autonomous with sandboxing [4] is Anthropic's own account of the sandbox idea behind Under the Hood.

REFERENCES

[1] 500+ Billion Tokens Later: Letting AI Agents Decompile A First-Person Shooter, Maurice Heumann, October 9, 2026, read October 19, 2026 (self-reported; token counts are estimates). ↑

[2] The Working Set of a Coding Agent, arXiv, August 17, 2026 (preprint). Not raised in the community chats; we added it because it bears on why a passing check can mask missing information. ↑

[3] Anthropic reinvents source control as Claude floods the world with code, The Register, October 13, 2026, read October 19, 2026 (trade press report). ↑

[4] Making Claude Code more secure and autonomous with sandboxing, Anthropic Engineering, read October 19, 2026 (vendor post). ↑

COMMUNITY INTEL

AI DevCon, November 2 to 4, Industry City, Brooklyn. Gareth will be at AI DevCon, and Tessl is offering our community 30% off with code ATV30. The event will be three days of context and skills engineering, evaluation, security and governance, software factories and team adoption, with 700+ senior engineers, 50+ talks and 10 workshops (the public programme page lists 700+ attendees and 52 sessions). Check out the details and get tickets.

ATV Worldis still taking pushes. Ahmed Abdelsalam wrote in the Community Call chat later that Friday that it is "multiplayer now." On the call, Shyam Sridhar promised to share "the prompt that was used"; the context was the game music, which his commit credits to Gemini. The live world was linked in the chat.

Hack & Furious, October 13: tooling access and adoption. David Crawford argued that innovation inside a large organisation tracks the tools people can access, and that the real adoption problem is workable tooling. Derek Williams's corollary, in his words: training for engineers amounts to access and opportunity, not classroom time (call discussion; no public link).

Carryover check. Issue 25 asked whether the promised ATVFD ontology demo (internal) had appeared, and Issue 27 asked whether the Graffiti Party claim graph (internal) had moved off manual OneDrive copying; we found no result for either. Issue 30's bet is still open until Issue 33. On October 12, Shyam Sridhar shared a deck of numbers in the Community Call chat; by Mark Rowe's reply it compared the Copilot app with the CLI, with a mention of "70% more tokens." David Crawford noted that VS Code holds two separate harnesses, and Mark asked whether the answers were the same quality, length and reasoning. A follow-up that answers that question would count. Our search covered the text of both chats, both calls and the three Teams channels for October 12 to 18; it could not see attachments or images.

Veni Vidi Vici, one of the ATV Team's channels, posted lessons 78 to 82 of "AI Engineering from Scratch" (internal, no public link), from Shyam Sridhar's account.

THE STACK

Secure AI-assisted development in VS Code: names enterprise policies for restricting MCP servers (ChatMCP) and disabling global auto-approval (ChatToolsAutoApprove). Its own caveat: its rule-based terminal auto-approval uses "best-effort command parsing that has known limitations."

CABRA, a synthetic benchmark: agents stay near-perfect on large tasks by offloading to tools like grep (a text search tool), but accuracy falls on the understanding-heavy task. Synthetic, so it does not size the cost on real repositories.

From Technical Debt to Cognitive and Intent Debt, Margaret-Anne Storey: the full paper behind the Workbench's term cognitive debt, "the erosion of shared understanding across a team." Reads as a conceptual paper, outside our 90-day window, kept as the origin of the terms.

Understanding the code is not enough, Bystams Bad Takes: the contrary view to the Getting Started habit. An unsigned opinion post, by a site whose contact is Fredrik Bystam, arguing design judgment matters more than line-level reading. No data.

SIGN OFF

This week's open question. David Crawford, Community Call, October 16: "How do you verify intent?" The nearest public data points are Anthropic's post on Scion, whose reviewers now examine the intent of a change and whose approval tooling separates comment-only changes from sensitive logic, and Heumann's byte check, where intent is the original program. We found no passage on where Scion's intent is written down for reviewers to check, and a byte check does not generalise. Bring a real attempt Friday.

Prediction (revisit Issue 33, sending November 3). At least one ATV-world Deploy run will carry a build-fallback warning ("Kept the live version of" or "didn't build, so it's under construction"). If so, the fallback caught something real. If every run stays green, the room's instant feedback is still the only gate we can see.

— The All The Vibes Team
Gareth Bland · David Crawford · Stephanie Schofield · Rajesh Singh · Shyam Sridhar · David Torres Bejarano

Issue 31 · Tuesday, October 20, 2026

Written with AI assistance.