Build your own harness

Start with the developer labs when you want to learn by changing a policy and inspecting the result. Work through tool recovery, context budgets, and approval across restart; then compose those contracts in the architecture builder and test your claims in the evidence workspace.

Continue with reliability, approval and restart.

The builder exports a versioned decision record for the design dossier. The evidence workspace keeps mock, offline, Codex functional, and live-provider claims distinct before you compare results.

Then compare designs in the evaluation workbench, inspect boundary response, choose security and operational boundaries, and use the design dossier to plan a separate harness implementation.

1. Read architecture.
2. Run the model/tool loop and its failure experiments.
3. Explore context, retrieval, and memory.
4. Exercise tool interoperability over MCP.

Start with architecture: a model proposes actions; the harness owns state, dispatch, authority, budgets, observation, and stopping.

The learning sequence is model boundary → tool loop → context → protocols → reliability → evaluation → execution/operations. The implementation backlog tracks each lesson. Examples are educational contracts, not a production SDK.

For existing components, use Choose. To interpret a result, use Evaluate.