AI Agent Trajectory Annotator and Reviewer

Type: Contract, hourly

Location: Remote

Hours: 20–30 per week

Pay: $20–30/hour, based on experience and language coverage

Start: Immediate

ABOUT THE ROLE

We evaluate how well advanced AI coding agents solve real engineering problems. An agent is given a real open source codebase inside a container and a hard task, then works on its own for 80 to 250 steps. A trajectory is the full record of that run — every command, result, and decision.

You will do two jobs, and you should expect either on any given day.

•Annotate — Read a trajectory nobody has looked at yet and judge it step by step.

• Review — Take an existing annotation, written by our AI tooling or another person, and confirm, correct, or reject it.

TASKS YOU'LL SEE

• Feature build — Add a working feature to a live library without breaking anything that already worked.

• Rebuild — Work out what a compiled tool does by running it, then rebuild it to match its output, exit codes, and file effects.

• Bug hunt — Find and fix twenty undocumented bugs across a dozen files with no test suite, then record what caused them.

Mostly Python and Go, with some Rust, C, and Ct+. A trajectory runs about 100 steps.

WHAT YOU JUDGE IN A TRAJECTORY

• Was the command right for the state the environment was actually in?

• Did the agent read the previous output correctly?

• Was the step wrong, or only inefficient — these are scored differently.

• Where did the run first go off course — usually earlier than where it visibly broke.

• Did the agent notice its own mistake and recover, or keep building on a false assumption?

• Did it game the grader instead of solving the task (e.g., weakening a test or hardcoding an expected value)?

WHAT WE NEED FROM YOU

• Experience — 2+ years in software engineering, DevOps, or site reliability, with real debugging in real codebases.

• Languages — Strong in Python or Go, and able to read a language you've never used.

• Linux — Comfortable with logs, running processes, build failures, and containers.

• Workflow — Everyday Git, diffs, pull requests, and issue tracking.

• Debugging — Able to work with no test suite and no error message pointing at the cause.

• Focus — Able to hold context across a long run, because step 74 can depend on step 12.

• Writing — Clear English, since every judgement needs an explanation another engineer can check.