The flagship path

Production agent systems.

Take one agent system from a written brief to a deployed system you have broken on purpose, and leave behind evidence a stranger can check.

9 stages. Each ends in an artifact a stranger can open and an eval that can be re-run. The order is the order production forces the decisions on you.

1. 01

   System brief

   Whose job is this and what would make you stop building it.

   Leaves: System brief2 eval checks

2. 02

   Architecture decision

   Which shape, and which credible alternative you are rejecting and why.

   Leaves: Architecture decision record2 eval checksindependent reviewlab

3. 03

   Tool and authority model

   What the agent may do, as what principal, and how you take it back.

   Leaves: Tool and authority model3 eval checkslab

4. 04

   Eval harness

   What a regression looks like, before you are allowed to ship a change.

   Leaves: Eval harness2 eval checkslab

5. 05

   Threat model

   Which inputs you do not trust, and which risks you are consciously accepting.

   Leaves: Threat model2 eval checkslab

6. 06

   Cost model

   What one unit of work costs and what stops the bill when it runs away.

   Leaves: Cost model2 eval checks

7. 07

   Deployment

   Put it somewhere a stranger can reach, with a rollback you have tested.

   Leaves: Deployment record2 eval checks

8. 08

   Incident simulation

   Break it on purpose and find out whether your instrumentation notices.

   Leaves: Incident simulation report3 eval checksindependent reviewlab

9. 09

   Public-safe portfolio proof

   What you can show publicly without leaking an employer or a customer.

   Leaves: Public-safe portfolio proof2 eval checks