A deterministic governance protocol that wraps probabilistic AI agents in 8 operational rules, a canonical JSON spec, and single-source-of-truth file discipline. No more blind writes. No more forked files. No more agents undoing each other’s work.
Long-horizon tasks exceed context windows. Agents compact, summarize, and silently drop critical constraints. By step 12, the agent that started your task is not the same agent finishing it.
Hand a task from Claude to Codex and each one subtly reshapes the objective. Without a canonical spec on disk, every agent reads a slightly different version of “what we’re building.”
Two agents touch the same file without coordination. One writes. The other overwrites. You lose 4 hours of work in a single commit — and don’t notice until production breaks.
of all long-horizon agent failures trace directly to context and goal drift — not reasoning errors, not hallucination. The agents are fine. The interface between them is broken.
Every task gets a 16-field JSON spec written to .agents/tasks/. It defines the objective, write scope, blast radius, validation commands, acceptance criteria, and rollback authority.
Single Source of Truth means there is exactly one canonical version of every active spec. Agents read it from disk. Agents write status updates back to the same file. No forks, no duplicates.
Eight human-readable governance rules that every agent must follow: re-read before act, version bump on every write, explicit write scope, one-way status transitions, no blind writes, second-agent review, blast-radius containment, and rollback authority.
Every agent re-reads the canonical spec from disk before executing. No relying on cached context or chat history.
Every write to a spec increments the version field. Stale reads are immediately detectable.
The spec declares exactly which files and directories the executing agent is allowed to touch.
Task status only moves forward: pending → active → review → done. No going backwards without a new spec.
An agent cannot write to a file it hasn’t read in the current session. Every edit requires fresh context.
No spec ships without a second agent validating the output against the acceptance criteria.
Every spec declares its blast radius — the maximum scope of files, services, and systems the task may affect.
Every spec names who can roll back, how, and under what conditions. Recovery is pre-documented.
The full WCS AI Triad™ package — schema, template, rules, folder structure, and live case study — starting at $249.