How to build an AI harness for the SDLC

Six steps to build an agent's working environment for the SDLC, from acceptance criteria and permissions to tests, persistent state, and outcome evaluation.

Contract
Define a small task, acceptance criteria, verification, scope, and an owner.
Context
Use AGENTS.md as the map to rules, architecture, and examples, and prepare a reproducible environment.
Tools
Connect tools with isolation, least privilege, argument validation, and a check against duplicate writes.
Loop
Plan, act, and check. Persist progress and evidence, with limits on time, cost, and retries.
Evidence
Require tests, a security check, and a review of the change. Match approval and rollback to the risk.
Evaluation
Compare results on representative tasks, and re-evaluate when the model, rules, or tools change.