2026 Anthropic Software Engineer Onsite PR Review Round, What It Tests: 2 Experiences
- Company
- Anthropic
- Reports
- 2 interview reports
- Level
- Unknown
- Stages
- Onsite
Anthropic's newest onsite round hands you someone else's code instead of a blank file: four pull requests in a repository you've never seen, with Claude available the whole time.
Coverage: Anthropic software engineer onsites in 2026: two candidates who took the round (a general SWE who was switching jobs, interviewed Jul–Sep 2026, and an infrastructure SWE, interviewed Apr–Jun 2026), plus what other candidates were told about it before their interviews. Both candidates were referrals and both were rejected after the onsite.
What the round is
Both candidates who took it describe the same setup:
- An unfamiliar codebase. A repository you haven't seen before. One candidate said it was something AI-related and thought each person might get a different one.
- Four open pull requests. One candidate described them sitting on a GitHub-like page next to an online VS Code-style editor.
- An AI assistant in the room. One candidate had Claude Code; the other used Claude to make the fixes.
- A language you don't pick. One was told TypeScript and couldn't switch. The other got a language they didn't know at all.
The official prompt one candidate received describes it as a test of writing and reviewing code with AI tools. It gives you the Claude Code CLI, lets you search Google and Stack Overflow, and puts you in a browser environment that looks like VS Code plus GitHub. Portals and recruiters have called it "agentic coding," "vibe coding," "slop review," "review PRs," and simply "code review."
What you're asked to do
The general SWE's version had two parts:
- Review every pull request. Work out what problem each one is trying to solve, then leave a review comment on it.
- Improve one. Pick a pull request, make it better, and push an update to it.
The infrastructure SWE's version leaned harder on fixing: they worked through the four pull requests and got two of them fixed with Claude before time ran out.
What it tests
Going by what these two candidates were asked to do:
- Reading code you didn't write. Both had to understand a new repository fast enough to say what each change was for.
- Judging a change. The first part is a review: what the pull request does, and what you'd say about it.
- Getting real edits done with an AI assistant. Claude is part of the setup, and the infrastructure SWE described their progress as pull requests fixed.
- Choosing what to spend time on. Time was tight. The infrastructure SWE left the pull request they understood least for last and never got to it.
Neither candidate learned how they were scored. The general SWE said it was hard to tell where the signal was when Claude Code could answer most direct questions about the code. Nobody reported whether running the code or tests was expected; that candidate didn't run anything.
Versions of the round
- General, infrastructure, full-stack/frontend and new grad SWE (2026): the pull-request round with Claude allowed; where a language was mentioned, it was TypeScript. A new grad was told AI could be used; a frontend candidate's portal hint said "review PRs"; a product-team candidate's hint said "slop review."
- Security Engineer (Sep–Nov 2026): a separate pull-request review focused on finding security issues. Two candidates were told no AI tools, and one got it as an extra round after a weak coding interview.
- Not the same round: at least six 2026 Anthropic onsite reports describe a system design round built on an existing design doc you critique, and some coding rounds start from given code with a bug in it. Those are different formats from the four-PR round.
Questions we found
AI Coding
| Question | Frequency | All last seen dates |
|---|---|---|
| Review four pull requests, then improve one (Claude Code allowed, TypeScript only) | 1 report | Jul–Sep 2026 (interview period; report date 2026-09-04) |
| Work through four pull requests in an unfamiliar repository and language, using Claude to make fixes | 1 report | Apr–Jun 2026 (interview period; report date 2026-08-12) |
System Design
| Question | Frequency | All last seen dates |
|---|---|---|
| Design a Model-Weight Deployment Platform | 1 report | Reports: Apr–Jun 2026 (interview period; report date 2026-08-12); Additional FastPrep history: 2026-08-21, 2026-08-13, 2026-07-29, 2026-05-15 |
| One-to-One Chat Service | 1 report | Apr–Jun 2026 (interview period; report date 2026-08-12) |
2 detailed candidate timelines
01: Four pull requests, Claude Code, and no choice of language
Report date: 2026-09-04 · Interview period: Jul–Sep 2026 · Role: Software Engineer, general hiring (not a Claude Code team role) · Level: experienced, switching jobs · Source: referral · Result: rejected after the onsite
Before the onsite
- The candidate portal's prompt named this round "agentic coding." It said Claude Code could be used and that the task was reviewing pull requests.
- The recruiter had also told them about the round ahead of time.
Onsite: the pull-request round
- The setup was an online VS Code-style editor plus a GitHub-like site holding four open pull requests. Claude Code was available.
- Part 1: go through every pull request, explain what problem each one solves, and leave a review comment on each.
- Part 2: pick one pull request, improve it, and push an update to it.
- The language was TypeScript. It wasn't their strongest language, and when they asked to switch, the interviewer said no.
- The repository was something AI-related. They thought each candidate might get a different repository.
- They didn't run the code, and they weren't sure whether running it was expected.
They came away unsure what the round was measuring, since Claude Code could answer most direct questions about the code. The other onsite rounds weren't described.
02: "Vibe coding" through four pull requests in a new repository
Report date: 2026-08-12 · Interview period: Apr–Jun 2026 · Role: Infrastructure Software Engineer · Level: not stated · Source: referral · Result: rejected; the onsite ended on a Friday and the rejection came by end of day the following Tuesday
Phone screen
- System design: a one-to-one chat service.
Onsite
- Vibe coding. A repository they had never seen, in a language they didn't know, with four pull requests to work through.
- They got two of the pull requests fixed. They read through a third but ran out of time before Claude could fix it. The fourth was in an area they didn't understand at all, so they put it last and spent their time on the three they could follow. Afterward they weren't sure that was the right call.
- System design: a model download service, matching Design a Model-Weight Deployment Platform.
- Hiring manager, culture, and past-project rounds, all described as standard.
What other candidates were told before the round
None of the people below had taken the round when they posted. Each line is one candidate describing their own invitation, portal hint, or recruiter note, newest first.
- Security Engineer, early November 2026. After a weak coding interview, the recruiter added another round for more signal: review a pull request for security issues. Claude and other AI tools were not allowed.
- Security Engineer track, September–October 2026. One candidate (11–15 years of experience) was told the next step was a project discussion plus a code review done on a pull request. A second security candidate was also scheduled for a pull-request review, and HR told them no AI tools could be used.
- Product-team candidate, October 2026. Their portal hint called the round "slop review."
- Frontend candidate, October 2026. Their portal hint said the round was about reviewing pull requests.
- New grad, General SWE, October 2026. Told the round was a code review in TypeScript, and that AI could be used.
- General SWE, October 2026. An HR email described an onsite "AI coding" round whose prompt asked them to fix bugs.
- Infrastructure SWE, Apr–Jun 2026 cycle (posted late June). Their onsite prompt said the round tests writing and reviewing code with AI tools, with the Claude Code CLI plus Google and Stack Overflow, in a browser environment that looks like VS Code and GitHub. Over July to September, four more candidates said they got the same or a very similar prompt; one of them said they weren't interviewing for infrastructure.
AI coding rounds at other companies
- Meta: Meta AI coding interview: two onsite rounds where you finish a hard algorithm inside a mostly written codebase with an AI model, then face stress-test suites.
- Google: Google AI coding interview: an online assessment with HackerRank's built-in assistant, and a live round on a multi-file project where using AI was optional.
- Stripe: Stripe AI coding interview: experienced engineers kept getting the same transaction rule parser, built step by step with a built-in model.
- Stripe interns: Stripe intern AI coding round: ten intern accounts of the same AI-assisted round.
- Rippling: Rippling AI coding interview: HackerRank's built-in AI on Rippling's familiar problems, plus an HR codebase at the onsite.
- DoorDash: DoorDash AI coding interview: AI CodeCraft: build and run several small services live with your own AI tool.
- LinkedIn: LinkedIn AI coding interview: one long, multi-part task in an IDE whose agent edits files directly.
- Shopify: Shopify AI coding interview: pair programming with your own AI on Shopify's well-known question set.
- Datadog: Datadog AI coding interview: bring any AI tool and share your screen; the tasks are easy, and the follow-ups decide the round.
- Capital One: Capital One AI coding interview: new grad Power Day agentic round: fix a bug, then add a feature to a large full-stack app with Claude Code.
- Bloomberg: Bloomberg AI coding interview: intern rounds that start in a small file and move to a large codebase, debugging with AI help.
- xAI: xAI AI coding interview: new grad take-homes and onsites where you build with AI.
- Adobe: Adobe AI coding interview: AI showed up as one step in a director round and in a take-home; the rest stayed classic.
- Snowflake: Snowflake AI coding interview: design and build a rate limiter or a retrying task queue while the AI writes the code; one AI Engineer onsite was in person with your own laptop.
- Snap: Snap AI coding interview: a backend-only chat service from an empty Python project with Claude or Codex, or a feature in an existing codebase.
- Ramp: Ramp AI coding interview: build inside a repo Ramp provides; one repo logged every prompt for the interviewer to read.
- Apple: Apple AI coding interview: CoderPad with a sidebar agent, alongside Apple's other recent interviews.
More companies with AI rounds on FastPrep:
- Amazon: Amazon interview report: an AI-assisted debugging OA (1 experience).
- Atlassian: Atlassian interview report: project coding and AI tool use (11 experiences, two parts).
- Circle: Circle interview report: three AI rounds: building, debugging and PR review (1 experience).
- Volta: Volta interview report: a deliberately vague data task with a token budget (1 experience).
Across companies: 2026 AI Coding OAs at 15 Companies: 32 Reported Problems and 2025–2026 AI Coding Phone Screens and Onsites at 43 Companies: 76 Reported Problems.
Sources: Anywhere useful interview info can be found.
Comments
0No comments yet — be the first to share what you know.