Coding Harness
An engineering harness for building software with AI coding agents without losing architectural control. It wraps GitHub's SpecKit spec-driven workflow (specify → plan → tasks → implement) with autonomous agents that drive each stage, and adds critic agents that gate every stage against documents the team owns: a constitution, an architecture doc, and test principles.
AI-assisted coding is fast, but code that looks right can quietly drift from your architecture, skip tests, or break conventions. Coding Harness keeps agents fast and keeps the codebase maintainable, by requiring a plan to pass an architecture review before tasks are generated, and requiring tests and implementation to pass a quality review before they're considered done. A human is always the one who merges to main.
- Test-first by design: failing tests are written before implementation code, and that's enforced by the workflow, not left to discipline
- Critic loops at every stage: plan, tasks, tests, and implementation each pass through an automated review before moving forward
- Grounded in project-owned documents, so the critics are checking against your team's actual standards, not generic best practice
- Supports a human-in-the-loop mode (review and approve each stage yourself) or a fully automatic mode (let the pipeline run end to end after you've reviewed the spec)
- Optional local LLM support for critic passes, if you'd rather not send every review through a hosted model