Anthropic Agent Coding Interview

Anthropic’s recruiter brief for this interview is explicit: “This interview will test writing code with LLMs as a building block, and prompting to create an agent. You should be conceptually familiar with Tool Use and Agent Loops and how they are handled in our API. The interview is open-book and we provide starter code for using the API, so there’s no need to memorize exact API syntax.”

In plain terms: it is not a LeetCode round. You get a working starter cell that wraps the Messages API and a tool-use schema, and your job is to build the agent looparound it — parse the model’s response, dispatch its tool_use calls to your code, feed tool_result blocks back, and iterate until the model is done.

The most concrete instance candidates describe is a stock-price agent. The starter code answers questions that need only a single tool call; you extend it into an agent loop that handles questions requiring multi-step reasoning or multiple tool calls. A second part asks you to reduce the number of turns the agent needs — part prompt engineering, part loop structuring.

A closely related three-part variant scores the progression explicitly, and this is the shape most recent reports describe:

1. Repair the supplied bugs until the agent works.
2. Make it answer the full question set within two turns.
3. Compress the workflow to one turn.

One candidate who completed all three parts in 55 minutes noted the interview version felt like the published one “with a slight variation” — so drilling the exact loop, not memorizing one answer, is what pays off.

• •~55 minutes, over Google Meet with screen-share, typically in a Colab notebook.
• •Starter cell wraps the Messages API and the tool-use schema — no need to memorize syntax.
• •Open-book on documentation; the interviewer watches how you iterate on prompts and structure the loop.
• •Appears in SWE phone screens (portal may show only “coding interview”, no question number) and on the research track (RE / RS / Fellowship).
• •The coding round is team-dependent — some candidates get a concurrency/systems prompt instead. Confirm the flavor with your recruiter.

• •Run through Anthropic’s public tool-use documentation once, end to end: define a tool, route tool_use turns back to your code, return tool_result blocks, and loop until end_turn.
• •Practice writing an agent loop from a blank notebook in under 20 minutes.
• •Keep a mental template for the prompt: role description, tool-catalog summary, output contract, exit condition — so you can refactor mid-round when the interviewer pushes on turn count or robustness.