<- all tokdocs

Reviewing AI-Written Pull Requests: Should Agent Traces Be Required?

By , founder of Mavric and Apso

View on TikTok ->

When an AI agent writes code, the pull request shows the final diff but hides the approaches the agent tried and rejected, the assumptions it made and the decisions behind the change. A TikTok from @agenticengineering argues that reviewers should look at the intent, the decision and the agent trace instead. The core idea is backed by a dated 2026 essay and by research on rejected agent PRs, but the video credits it to a "Justin Bennett" essay that I could not find.

The saved copy of the video had no audio or transcript, so the quotes below are its on-screen captions.

What is the problem with AI-written pull requests?

The video's top banner reads "Your PR may be hiding the most important part of the work". Its captions describe an agent that "MAY HAVE BROKEN CODE", went through "AN AGENT REJECTING APPROACHES," and produced a PR where "THE REVIEWER SEES ALMOST NONE" of that work.

Graham Brooks makes the same argument in "Code Review After the Pull Request", dated 26 April 2026. In his account, the reasoning behind agent-written code sits in the prompt, the conversation and the tool calls. A pull request compresses that trail into a summary written after the fact.

His example is a commit labeled "fix flaky test" that hides an agent adding a sleep after several failed attempts. The diff looks small and harmless. The history behind it would tell a reviewer the test is still broken.

Who wrote the essay the video cites?

The caption reads "SO JUSTIN BENNETT WROTE AN INTERESTING", followed by "MUCH OF THE THINKING NOW LIVES" and "AND ARGUES THAT WE MAY BE". The video gives no title, link or handle, and the description lists only hashtags.

I found no author by that name writing about pull requests, agent traces or reviewing intent. The nearest match was Jim Bennett, a developer experience engineer at Arize AI, whose search results tied him to agent observability and a general post on reviewing pull requests. I did not open that page. The captions look auto-generated, so the name may be a transcription error. Either way, the attribution cannot be checked.

What should reviewers look at instead?

The captions ask "WHAT SHOULD BECOME REQUIRED?" and list options ending in "THE AGENT TRACE, OR ALL THREE?". They propose moving "TO REVIEWING THE INTENT, THE DECISION," along with "IMPORTANT ASSUMPTIONS". The speaker frames this as an open question, with "I THINK", so it is a proposal rather than a claim.

Brooks recommends a partly different set:

  • intent and specifications
  • system state over time
  • observed behavior
  • pre-review by other agents
  • sampled human audits, with the agent's logs consulted when something breaks

Both sources agree the reasoning matters. They differ on whether a person should read a trace on every PR. The video's closing caption, "AND THEN LET THE REVIEWER FOCUS", suggests narrowing what the reviewer reads, which fits Brooks's sampling approach better than a rule that every trace gets read.

Why do agent pull requests get rejected?

The MSR '26 preprint "Why Are Agentic Pull Requests Merged or Rejected? An Empirical Study" (arXiv 2605.22534v1, 21 May 2026) started with 11,048 closed agentic PRs from repositories with at least 500 stars. It kept 9,799 after removing bot-only reviews and manually inspected 717, of which 353 were rejected. Of those 353:

  • 35.7% (126) were clear agent failures
  • 31.2% (110) were workflow constraints such as duplicates or inactivity
  • 33.1% (117) showed no observable decision rationale, often because the PR was closed without a comment

That last figure is about reviewers' reasons for rejecting a PR, which is a different subject from the agent's rejected approaches. It still shows that in a third of sampled rejections, nobody can tell from the PR page why something happened. The study does not test whether a structured trace would fix that.

What the video does not show

The video is a talking head with captions. It shows no terminal, no agent trace, no PR template and no tool run. The hashtags include #codex and #claude, but no caption names either product.

Bottom line

The argument holds up: a diff from an AI agent hides the attempts and assumptions that produced it, and Brooks's essay makes the same case with a concrete example. What to require instead is still open, with options ranging from reviewing intent to sampled audits of agent logs. The "Justin Bennett" essay the video cites could not be found, so credit the idea to Brooks's published version until a source turns up.

Sources

Published October 29, 2026. Based on a TikTok by @agenticengineering. Full video notes (markdown)