Module 07 Activity

Scenario

A tool or two, and the question of whether you needed an agent at all.

What you build

Tools with model-facing descriptions and real validation, an enforced registry, bounded stopping conditions, and argument-level logging.

Steps

1. Write one tool and check what the model actually sees: name, description, argument schema. Rewrite the docstring for the model.
2. Put the guarantees in validation rather than the docstring, and test every refusal branch directly.
3. Enforce a tool registry and include the real names in every rejection. Measure your invention rate.
4. Implement all four stopping conditions and show which fires first on a tight loop.
5. Log tool arguments as well as names, and find one wasted call caused by near-duplicate arguments.
6. For each task in your system, decide chain or agent using the run-time-sequence test.

Evidence to hand in

• What the model sees for one tool, before and after rewriting.
• Direct tests of every refusal branch.
• The registry, a rejection message, and your invention rate.
• The four stopping conditions with which fired first.
• A wasted call found from argument logging.
• The chain-or-agent decision per task.

Review checklist

• The docstring is written for the model, not for developers.
• Guarantees live in validation, not prose.
• Rejections include the real tool names.
• All four stopping conditions exist.
• Arguments are logged, not just tool names.
• Anything with a known sequence is a chain.

Mark complete