Agents / Coding with LLMs

Research-track 1-of-N option: write code in Colab that uses an LLM (via the Anthropic API) as a building block — designing prompts and a tool-use loop to solve a multi-step problem. Open-book on documentation; starter API code is supplied so you don't memorize syntax. Requirements The recruiter blurb 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 Agents 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." Reported shape: Colab notebook, 55 minutes on Google Meet, share-screen. Starter cell wraps the Messages API and the tool-use schema. The task itself is open-ended — design prompts for an agent that plans a small workflow, defines its own tools, and demonstrates progress on a concrete input. Specific tasks have not been described in detail publicly. Reported task — stock-price agent One concrete instance: the starter code answers stock-price questions that need only a single tool call, and you extend it into an agent loop that handles questions requiring multi-step reasoning / multiple tool calls. A second part asks you to reduce the number of turns the agent needs to reach an answer — partly a prompt-engineering exercise, partly a test of how you structure the loop. This round now also appears in standard SWE phone screens, not just the research-track menu: the portal may list only "coding interview" with no Q-number, and the recruiter blurb names Tool Use and Agent Loops as the prep target. Notes Open-book on docs, no AI assistants beyond the supplied API. Screen-share lets the interviewer watch how you iterate on prompts. Publicly available detail is thin because the round only goes to research-adjacent candidates who self-select it. Treat this card as a structural overview, and re-check forum closer to the interview date for fresh task-specific writeups. AI-assisted coding variant (Claude Code CLI) A distinct rotation reframes this round from building an agent to coding with one. The recruiter blurb: "This interview will evaluate your ability to write and review code using AI tools. You'll have access to the Claude Code CLI, and are allowed [to] search things up on Google, Stack Overflow, etc. You'll work on some practical engineering tasks in a browser-based environment that is similar to Visual Studio Code + GitHub. No account setup is required in advance." Here AI assistance is the point, not a banned tool — you drive the Claude Code CLI to complete and review practical engineering tasks in a hosted VS Code-like environment, and you are graded on how effectively you delegate, verify, and review AI-written code. This contrasts sharply with the standard live-coding rounds where assistants are forbidden. Task specifics are not public yet; treat it as a code-review plus agentic-workflow exercise rather than a from-scratch algorithm round, and practice reading and correcting AI-generated diffs quickly. Preparation Run through Anthropic's public tool-use cookbook end-to-end at least once: define a tool, route the assistant's tool_use turns back to your code, return tool_result blocks, iterate. Practice writing an agent loop from a blank Colab in under 20 minutes: parse the response, dispatch tool calls, stop on end_turn. Have a mental template for prompt structure: role description, tool catalog summary, output contract, exit condition. Be ready to refactor mid-round when the interviewer pushes on robustness or eval criteria.