Shokunin Harness
A governed environment for building software with AI coding agents.
The problem
Long agent sessions can lose relevant context, introduce unverified changes and report success without reproducible evidence. Longer prompts alone do not solve those operational failure modes.
What I am building
I am designing an agent-development harness with bounded responsibilities, typed handoffs, deterministic checks and an explicit evidence trail.
Decisions that shaped the work
Treat agent output as untrusted until independently checked.
Separate agent work, validation and handoff into explicit contracts.
Audit the guards themselves: a gate that says PASS without testing anything is not a gate.
Evidence and references
In development. The uploaded September 2026 audit reports fixes and test results for specific subsystems; it does not prove that the complete harness prevents hallucinations or is production-ready.