Agentic pull requests
Confidence 0.85 · 13 sources · last confirmed 2026-10-03
Pull requests authored by AI coding agents rather than by humans — and the review, merge and governance machinery that has grown up around them. As of early 2026 this is measurable at population scale, which makes it one of the few areas in this corpus where the practitioner claims can be checked against data.
Scale
2026-02-09-li-aidev-studying-ai-coding-agents-on-github (the AIDev dataset) is the reference population:
| Agent-authored PRs | 932,791 |
| Repositories | 116,211 |
| Developers involved | 72,189 |
| Agents | 5 — OpenAI Codex, Devin, GitHub Copilot, Cursor, Claude Code |
Nearly a million agent-authored PRs across 116k repositories is an established practice, not an emerging one. The 72,189 developers are the other half of the story: these land in front of humans who must do something about them.
The bottleneck moved
Once agents author most changes, review capacity — not authoring capacity — is the constraint. 2026-04-13-branco-lgtm-auto-merged-llm-agentic-prs opens on it: “AI tools are generating code faster than humans can properly review it, leading repositories to skip review and auto-merge agentic PRs directly.” Claire Vo built a bot for exactly this reason; Carson ships ~40 PRs a day with “no QA team reviewing a single one.” DORA sees the consequence at survey scale: throughput’s relationship with AI adoption turned positive in 2025 while delivery stability’s remained negative — volume is no longer the problem; the delivery system’s absorptive capacity is.
Demaree (Onebrief CEO, YC Root Access, Sept 2026) names the constraint that came after authoring at one company. Once teams were “good enough at delivering high-quality features that reliably work, they’re now delivered so fast that the DevSecOps pipeline basically breaks because you’re shipping PRs that are, you know, 180,000 lines of code.” His general rule is this section’s: “each time you unblock something, you earn the right to find out what the next blocker is.” The context is a reported 45x rise in feature delivery across Onebrief’s best teams since January 2026, a self-report (see ai-coding-productivity-evidence). A single 180,000-line PR is a different regime from the small, focused PRs that Branco et al. find auto-merged in public repositories, and the talk does not say how such PRs are reviewed or how the pipeline was fixed.
What actually happens to them
Rejection is close to a coin flip. 2026-06-11-abujadallah-rejection-of-agentic-pull-request-fixes: 46.41% of fixes proposed by Copilot, Devin, Cursor and Claude are rejected, in 14 reasons across 4 categories — incorrect implementation, CI/test failure, agent unable to implement, and low priority. That fourth category is not an agent failure at all: the fix was fine and nobody wanted it. Rejected work is not free — it consumes reviewer attention, CI capacity and tokens.
What merges without review is small and additive. 2026-04-13-branco-lgtm-auto-merged-llm-agentic-prs: auto-merged agentic PRs are smaller and more focused; maintainers are cautious toward PRs that delete code; OpenAI Codex and Claude Code get the highest auto-merge rates; and mature repositories more often reject auto-merge policies altogether.
Governance is bimodal, and that is the most consequential finding. “Repositories tend to either auto-merge all agentic PRs or none.” Auto-merge is a repository-level posture, not a per-change judgement. The graduated middle path that Merge Mommy and Land PR represent is rare in the wild — which makes the practitioner material more valuable as guidance and less valid as description.
What survives review still carries debt. 2026-03-30-liu-debt-behind-the-ai-boom: across 302.6k AI-authored commits in 6,299 repositories, 484,366 issues introduced, 89.3% code smells, >15% of commits from every assistant introducing at least one issue, and 22.7% still present at the latest revision.
The emerging gate design
The practitioner pattern that recurs is risk-scored partial autonomy: score each change, auto-approve a low band, escalate the rest to a human. Merge Mommy’s six dimensions are the most explicit statement of the axes —
blast radius · reversibility · data security · ops impact · verification gap · change surface
Two of these have independent support. Reversibility and external visibility are what 2026-05-14-pochampally-assistant-or-actor-delegation-regret finds actually drive humans to withdraw trust from an agent (more than stakes do), and both are what Singapore’s framework would place a human checkpoint on. Verification gap is the practitioner name for the quantity SpecBench formalises and shows growing 28 points per tenfold increase in code size.
Carson’s Land PR is the same idea with a different gate: fresh agent review, up to two loops, then a narrated video walkthrough with a red/green test list that he watches before approving the merge. That is preview, which the delegation-regret study identifies as the specific thing whose absence produces regret even on successful outputs.
What would make agentic PRs land better
The rejection study’s prescriptions: hint the approach, state forbidden approaches, instruct on CI validation without breaking changes, and prioritise properly. The first two align with 2026-02-12-gloaguen-evaluating-agents-md-repository-level-context-files, which finds that instructions in context files are followed while repository overviews are not helpful and cost >20% more inference. Specific procedural constraint earns its tokens; general background does not.
Related concepts
agentic-engineering, agent-fleet-management, agent-oversight-and-delegation, ai-generated-code-quality, reward-hacking, ai-coding-productivity-evidence.
Debates and supersession
- Is 46.41% rejection bad? It is a rejection rate among opened PRs, not a defect rate, and the “low priority” category shows some rejections are targeting failures rather than quality failures. Against that, human PR rejection rates in comparable open-source projects are not reported in any source here, so there is no baseline — the number is alarming without being interpretable. Open.
- Do practitioner gates generalise? 2026-04-13-branco-lgtm-auto-merged-llm-agentic-prs finds the wild is bimodal and mature repositories decline auto-merge. Neither Merge Mommy nor Land PR reports a false-approve rate, which is the only number that would establish whether the low-risk band is actually low-risk. Open.
- All population data is public GitHub, five agents, self-identified attribution. Enterprise and private-repo behaviour is entirely unobserved.
Two opposite conventions, in one episode (added 2026-09-16)
The Agent Factory, July 2026 puts two practices for agent-authored PRs side by side, twenty minutes apart, without noticing that they point in opposite directions.
Make authorship conspicuous. YK Sugi’s convention: have the agent open the PR as a draft, review it, then mark it ready.
“You can just ask the agent to create a draft PR… and then check it before marking it as ready for review so that people will [know] OK, this PR maybe was created by the agent. It’s not ready for review yet.”
This is an out-of-band provenance marker built from existing GitHub semantics — no new tooling, no bot, no label taxonomy. Given that this page’s central finding is that reviewer attention is the scarce resource, a zero-cost signal that says “do not spend attention here yet” is a cheaper intervention than anything else the corpus records.
Make authorship invisible. Lydia Hallie describes the opposite as standard practice inside Anthropic — the Claude GitHub app installed across their repositories, so that
“if your CI fails or someone leaves a comment, Claude will automatically try to fix it until your CI is green.”
with the stated goal “to be as out of the loop as possible.” A teammate’s review comment becomes an input to an automated fix loop rather than a request to a human author.
The conflict is about what green CI certifies. Making a passing pipeline the loop’s terminal state assumes CI failure is the failure mode. Abujadallah et al. find 46.41% of agent-proposed fixes rejected across four categories, only one of which is CI/test failure — incorrect implementation and low priority are both invisible to a green pipeline, and this page’s existing contradicts edge to reward-hacking says a passing suite is precisely what an optimising agent produces either way. An auto-fix loop that halts on green therefore halts on the signal least able to distinguish a good fix from a plausible one.
The reconciliation neither speaker offers is the one agent-oversight-and-delegation has converged on: calibrate to consequence and reversibility. Draft PRs for changes that reach production; auto-fix for CI-green plumbing. Both conventions are defensible inside their band and neither is defensible as a default.
CI speed as the ceiling on agent PR throughput (added 2026-09-18)
Nystrom and Vo (How I AI, May 2026) put the constraint upstream of review. An agent facing an hour-long CI run “is just going to sit there and spin”; at three minutes, a human and “your little swarm of agents” get far more done. Vo cites Stripe’s roughly 1,300 agent PRs a week — “you like cannot do that if your CI is slow” — and states it as a rule: there is “a true mathematical limit on your capacity to ship code to production that is a reflection of how fast your CI pipeline is.” Notion’s response is a project (Afterburner) to cut CI time to a quarter. The episode also shows an agent PR produced from a four-sentence task comment in about ten minutes, and the reviewer’s reply to a CI type error was simply “I don’t get it.”
AI review on every merge request, at company scale (added 2026-09-19)
The review-capacity bottleneck this page describes has a large-scale response on record. At Cloudflare, per Cloudflare Investor Day 2026, every merge request on the standard CI pipeline gets an AI review. Up to seven specialised agents (security, performance, code quality, documentation, compliance and others) run in parallel, and “a top-tier model” coordinates them and posts a single structured review. The first 30 days:
| Reviews | Merge requests | Repos | Median time | Median cost | Coverage |
|---|---|---|---|---|---|
| 131,246 | 48,095 | 5,169 | 3 min 39 s | $0.98 | 100% |
Two things this adds. First, a unit cost for machine review, under a dollar per review, which puts the economics of the gate on the page for the first time. Second, the architecture: specialist reviewers per concern, merged by a coordinator, which is the multi-dimension scoring of Merge Mommy run as separate agents. It is the same answer Adobe’s security leader reached from the other direction in Alim’s talk. Loading every team’s guidance into the authoring agent crowds its context, so checking each concern in a separate reviewing agent keeps it out of the authoring agent’s context.
What the deck does not report is the outcome: how many reviews found something real, how many were dismissed, and whether change failure rates moved. Coverage and cost describe the gate, not what it catches.
One firm’s agent share of authoring and review (added 2026-09-28)
Geoff Charles, Ramp CPO, at Lenny and Friends Summit (Sept 2026) gives both halves of the pipeline at one company. Authoring: Ramp’s in-house coding agent Inspect builds “75% of our PRs”, with about a thousand PRs in a month “submitted by a non-engineer”. Review: a review agent, Review Buddy, reads the codebase, the quality and security checks and “all the prompts that led you to build that code”, and “93% of our PRs are now automatically handled”, so senior engineers spend their time on the remaining 7%. The talk does not define “handled” (approved, commented or routed), and the figures are self-reported. The structure matches the “bottleneck moved” argument above: once authoring was automated, review was automated next, and testing (a QA agent) after that.