Downbeat

Downbeat turns long-running agentic work into an operable product. From the UI, an operator launches a real Goal → Work Package → Run journey. Specialist agents can work in parallel while the Rust kernel owns durable identity, authority boundaries, fencing, recovery, and proof joins.

For the Build Week demo, Downbeat launched six GPT-5.6 specialist roles over a real repository: legibility, operability, identity, authority, recovery, and proof. A final synthesizer joined their typed outputs into a tangible judge-ready report with exact source paths, one supported true claim, one rejected false claim, and explicit gaps. The run used 10 real Codex agent calls and preserved interruption/retry state rather than hiding it.

Codex accelerated the project throughout architecture exploration, Rust implementation, UI iteration, testing, packaging, installation, and the live demo. GPT-5.6 powered both the orchestration work and the specialist run shown in the video. The central product decision was to keep judgment in a persistent Codex root while Rust enforces the state and authority boundaries that should not depend on model declarations.

The source repository includes setup and product documentation. The attached Apple Silicon package is the exact build used for the demo.

Built With

• codex
• css
• gpt-5.6
• html
• javascript
• macos
• rust