---
title: "Reviewing AI-Written Pull Requests: Should Agent Traces Be Required?"
date: "2026-10-29"
video_id: "7702074482118118670"
tiktok_url: "https://www.tiktok.com/@agenticengineering/video/7702074482118118670"
creator: "@agenticengineering"
category: "Coding"
tags: ["code review", "pull requests", "ai coding agents", "agent trace", "agentic engineering"]
excerpt: "AI-written pull requests hide the agent's reasoning from reviewers. What a 2026 essay and an MSR '26 study say about reviewing intent and agent traces instead."
---
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](https://www.grahambrooks.com/post/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](https://arxiv.org/html/2605.22534v1)" (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

- [Code Review After the Pull Request, Graham Brooks, 26 April 2026](https://www.grahambrooks.com/post/code-review-after-the-pull-request/): reasoning outside the diff, the flaky-test example and an alternative review focus.
- [Why Are Agentic Pull Requests Merged or Rejected? An Empirical Study, arXiv 2605.22534v1](https://arxiv.org/html/2605.22534v1): dataset sizes and the rejection breakdown.
- [TikTok post by @agenticengineering](https://www.tiktok.com/@agenticengineering/video/7702074482118118670): the video and captions reviewed here.


## Video description and verification notes

**The file is 124.67 s (ffprobe duration 124.666667), 1080x1920, H.265/HEVC Main (hvc1) in an MP4/MOV container, 30 fps. The video stream runs at 633,294 bps and the container at 550,411 bps. The file is 8,577,246 bytes and has a single video stream with no audio stream, although metadata.json lists acodec aac and format id bytevc1_1080p_635337-1. The post timestamp is 1793278974, which is 2026-10-29 in UTC. The handle is @agenticengineering (uploader_id 7591537776035939342), the nickname is "Agentic Engineering", and the sound credit is "original sound" by Agentic Engineering. The description is 119 characters: "Your PR may be hiding the most important part of the work #ai #agenticengineering #vibecoding #codex #claude". At capture (2026-10-29T20:01:50Z) the video had 377 views, 20 likes, 7 comments, 3 reposts and 5 saves. No transcript.txt, transcript.json or .srt file was readable in the staged directory, and the downloaded file has no audio, so I have no spoken-word transcript and no segment timestamps. A missing audio stream in the download does not prove the original TikTok is silent. The evidence below is the burned-in captions alone, from the 17 frames I read: 1, 5, 10, 15, 18, 20, 22, 25, 27, 30, 35, 40, 45, 50, 55, 60 and 62. These frames fall at roughly 2-second spacing, so frame N is near (N-1)×2 s, and I have not read every frame. Every frame shows one man (the sound credit and handle suggest the channel owner, but no frame names him) in a black "CODE BLOODED" t-shirt, seated in a room with a marbled painting on the left and a recessed white shelf on the right. A laptop screen is partly visible at the lower right edge from frame 10 onward and its contents are unreadable. A blue box at the top of frames 1 and 5 reads "Your PR may be hiding the most important part of the work". Word-by-word captions at the bottom read, in the order of the frames I read: "A I. MAY HAVE BROKEN CODE" (frame 1), "AN AGENT REJECTING APPROACHES," (frame 5), "AND THE REVIEWER SEES ALMOST NONE" (frame 10), "SO THE REAL WORKFLOW BECOMES HUMAN" (frame 15), "SO JUSTIN BENNETT WROTE AN INTERESTING" (frame 18), "MUCH OF THE THINKING NOW LIVES" (frame 20), "OF THE ENGINEER'S THINKING." (frame 22), "AND ARGUES THAT WE MAY BE" (frame 25), "AND WHERE THINGS CHANGED. I THINK" (frame 27), "R, WHAT SHOULD BECOME REQUIRED?" (frame 30), "THE AGENT TRACE, OR ALL THREE?" (frame 35), "STRUCTURED SOMETHING." (frame 40), "IMPORTANT ASSUMPTIONS. YOU DON'T NEED TO" (frame 45), "TO REVIEWING THE INTENT, THE DECISION," (frame 50), "WITH THE TWO HUMANS BARELY SEEING" (frame 55), and "AND THEN LET THE REVIEWER FOCUS" (frame 60). Frame 62 is at about 122 s. No frame shows a terminal, a PR page, a repository, a URL, an article, a slide or a chart; the video is a talking head with captions only. The captions are fragments, and the frame numbers I read are not consecutive, so I do not reconstruct a full script. The argument the captions suggest is that AI-assisted PRs show reviewers a diff but hide the agent's rejected approaches, assumptions and decisions, so the author wants the intent, the decision and the agent trace to be reviewed. Because there is no transcript, the check of transcript end against audio duration cannot be performed.**

## The attribution to "Justin Bennett" could not be sourced

The caption in frame 18 (about 34 s) says "SO JUSTIN BENNETT WROTE AN INTERESTING", and the captions in frames 20 to 25 continue with "MUCH OF THE THINKING NOW LIVES" and "AND ARGUES THAT WE MAY BE". The video therefore presents a named writer's essay as the origin of its argument. I searched for a Justin Bennett who wrote about pull requests, agent traces or reviewing intent and found no such author or article. The nearest name that surfaced was Jim Bennett, a developer experience engineer at Arize AI, but the results I saw tied him to agent observability and to a general post on reviewing pull requests, not to the claim in this video. I did not open that page, so I make no claim about its content. The video gives no URL, handle or title for the essay, and the description lists only hashtags. I could not follow the pointer, so the attribution stays unverified. The name may also be a caption transcription error, since the captions look auto-generated. I therefore do not assert that a person named Justin Bennett does or does not exist.

## The core claim, that a PR hides the agent's reasoning, matches other 2026 writing

The caption sequence in frames 1 to 15 says the AI may have broken code, rejected approaches along the way, and that the reviewer "SEES ALMOST NONE" of it. Graham Brooks's essay "Code Review After the Pull Request", dated 26 April 2026, argues the same structure. In Brooks's account, the reasoning behind agent-written code sits in the prompt, the conversation and the tool calls, and a pull request compresses that trail into a summary written after the fact. His example is a commit labeled "fix flaky test" that conceals an agent adding a sleep after several failed attempts. This is a different author from the one the video names, so it does not confirm the video's attribution. It does show that the idea the captions describe has a published, dated version.

## "Review the intent, the decision, and the agent trace" is a proposal, and the sources differ on what to review

At frames 30 to 50 (about 58 to 98 s) the captions ask "WHAT SHOULD BECOME REQUIRED?", list "THE AGENT TRACE, OR ALL THREE?" and say reviewers should move "TO REVIEWING THE INTENT, THE DECISION," with "IMPORTANT ASSUMPTIONS" and "STRUCTURED SOMETHING." Because the sentences are fragments, I cannot tell what the three items are beyond the intent, the decision and the agent trace. The video frames this as an open question ("I THINK" at frame 27), so it makes no factual claim I can mark true or false. Brooks recommends a partly different set. He lists intent and specifications, system state over time, observed behavior, pre-review by other agents, and sampled human audits, with the agent's logs consulted when something breaks. So the sources agree that the reasoning matters and differ on whether a human should read a trace on every PR. The closing caption in frame 60 (about 118 s), "AND THEN LET THE REVIEWER FOCUS", suggests the point is to narrow what the reviewer must read, which fits Brooks's sampling approach more than a rule that every trace be read.

## Evidence on why agent PRs get rejected shows the missing-rationale problem in data

Frame 5 (about 8 s) has "AN AGENT REJECTING APPROACHES". A related finding is in the MSR '26 preprint "Why Are Agentic Pull Requests Merged or Rejected? An Empirical Study" (arXiv 2605.22534v1, 21 May 2026). It began with 11,048 closed agentic PRs from repositories with at least 500 stars, 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, and 33.1% (117) showed no observable decision rationale, often because the PR was closed without a comment. That last figure concerns the reasons for rejecting a PR, which is a different subject from the agent's rejected approaches that the video describes. It does show that decision rationale is absent from a third of the sampled rejections, so a reader often cannot reconstruct why something happened from the PR page alone. It does not test the video's proposal that a structured trace would fix this.

## What the video does not show

The video shows no terminal output, no agent trace, no sample PR template, and no run of any tool. The hashtags #codex and #claude appear in the description, but no frame I read names Codex, Claude or any other product, so I make no product claim. The only named entity is the writer in frame 18, and the only quantity is "TWO HUMANS" in frame 55 (about 108 s), where the caption "WITH THE TWO HUMANS BARELY SEEING" appears. I read it as the author and the reviewer, but the words before it are missing from the frames I read, so I do not assert that interpretation.

## Key Takeaways

- **Unverified:** The video says at about 34 s (frame 18) that "Justin Bennett wrote an interesting" piece. I found no such article or author, and the video gives no link.
- **Partial correction:** The video presents the argument as coming from one named writer. A dated essay by a different author, Graham Brooks (26 April 2026), makes the same core point, so the idea does not trace only to the named source.
- **Verified:** The claim that a PR hides the agent's reasoning matches Brooks's description of reasoning living in prompts, conversations and tool calls rather than in the diff.
- **Context:** In the MSR '26 preprint, 33.1% of 353 manually inspected rejected agentic PRs had no observable decision rationale, which concerns reviewers' reasons, not the agent's.
- **Clarification:** No transcript or audio exists in the staged file, so every quotation here is a burned-in caption. Frame timestamps are approximate and the captions are fragments.

## Resources

- [Why Are Agentic Pull Requests Merged or Rejected? An Empirical Study, arXiv 2605.22534v1, 21 May 2026](https://arxiv.org/html/2605.22534v1) establishes the dataset sizes (11,048, 9,799, 717 and 353) and the 35.7%, 31.2% and 33.1% rejection breakdown I cite.
- [Code Review After the Pull Request, Graham Brooks, 26 April 2026](https://www.grahambrooks.com/post/code-review-after-the-pull-request/) establishes the argument that agent reasoning lives outside the diff and the alternative review focus I compare against.
