Governed agent design workshop

Design-only roadmap exercise

Design a bounded agent that prepares a change request, reads an approved inventory, proposes an update, obtains human approval, and submits to a synthetic ticket tool. Do not connect it to production systems.

Tasks

1. Define purpose, prohibited actions, actors, resources, data, autonomy, budgets, and stop conditions.
2. Draw identity, trust, tool, memory, approval, execution, and evidence boundaries.
3. Build the allowlisted tool registry and typed contracts.
4. Define per-action authorization and bounded delegation.
5. Create approval screens for the material side effect and changed-plan handling.
6. Specify workspace, network, secret, step, time, and cost confinement.
7. Design the run receipt and incident response.
8. Create evaluation cases for injection, malicious tool results, revoked access, partial success, retry, and cancellation.

Deliverables

Submit the authority model, boundary diagram, tool registry, policy table, approval contract, confinement profile, memory policy, receipt schema, evaluation set, and incident playbook.

This exercise evaluates architecture completeness only; it is not the unreleased executable lab or practical certification.