Specification-driven development for AI coding agents
Agents forget what they said three steps ago. No persistent memory of intent.
Code gets written but never verified against the original intent.
One agent, one task, one branch — even when the work is independent.
A single pass produces a rough draft, not production code.
Define the what
Plan the order
Run the loop
Verify the result
Blueprints are the source of truth. Agents read them, build from them, and validate against them. When something breaks, the system traces the failure back to the blueprint — not the code.
Claude builds. Codex challenges. A second model with different training biases catches what the first one missed.
| LEVEL | DESCRIPTION | ACTION |
|---|---|---|
| P0 | Security / data loss | Block merge |
| P1 | Spec violation | Block tier advance |
| P2 | Quality concern | Flag for review |
| P3 | Suggestion | Note only |
All Codex features are additive. Blueprint works fully without Codex installed.
Each iteration validates against the blueprint. The loop runs until all tasks pass or the iteration limit is reached.
Circuit breakers isolate failures to individual agents. One broken task does not block the rest of the wave.
Requires Claude Code, git, macOS/Linux. Optional: tmux for parallel agents.