Lab: A single-purpose agent

Build a single-purpose agent that solves a specific task you care about. Two to five tools, bounded iteration, structured logging, and an eval set. Submit the working code and a written analysis of its behavior on the eval set.

Deliverables

• — The task specification and success criteria.
• — Working agent code.
• — Tool specifications with descriptions.
• — An eval set of at least 15 realistic scenarios.
• — Structured logs from running the eval set.
• — A one-page analysis of the pass rate and the failure modes.

Rubric

• — The agent completes the task on at least 60% of the eval set.
• — Iteration is bounded and logged.
• — Failure modes are named and traced to specific causes.
• — The code is under 500 lines; if longer, the extra length is justified.