Module 04 Activity

Scenario

Your agents are about to be given tools, knowledge and access. Everything here is about making misuse impossible rather than discouraged.

What you build

A tool layer that is read-only where possible, self-validating always, and tested without any agent - plus per-agent scoping and staleness budgets.

Steps

1. Write one tool with all three safety properties: read-only, validates its own input, never raises.
2. Test it directly with .run() against every refusal branch, including a hostile input. No agent, no key.
3. Scope tools per agent so that only one agent holds any write capability. Print each agent's tool list.
4. Give each knowledge source a staleness budget based on how often that kind of document changes, and flag anything past it.
5. Take a multi-turn exchange and extract the established facts into named fields instead of passing the transcript.
6. Report your tool test coverage as a fraction of branches, not of tools.

Evidence to hand in

• The tool with its three properties identified.
• Direct test results covering every refusal branch.
• Per-agent tool lists showing where write access lives.
• The staleness budgets with the change-frequency reasoning.
• The extracted-fields version of a conversation.
• Branch coverage as a fraction.

Review checklist

• The tool returns strings on failure and never raises.
• Every refusal branch has a direct test.
• Exactly one agent holds write access, and the others physically lack it.
• Staleness budgets differ per source with a stated reason.
• The conversation was reduced to fields, not summarised into prose.

Mark complete