Agent Council is an experimental open-core project for evaluating whether blind, heterogeneous AI-agent deliberation can produce better technical and product decisions than strong single-agent baselines.
The v0 implementation is intentionally small:
- Claude Code CLI
- Codex CLI
- subscription-first execution
- blind independent research
- controlled cross-review
- adversarial challenge and rebuttal
- dual independent judges
- Visibility Firewall enforced at the filesystem boundary
- reproducible evaluation against single-agent baselines
Pre-v0 / validation-first.
The governed Phase H technical value-validation pilot completed on September 3, 2026. It evaluated the frozen 10-case technical corpus with 120 evaluator calls and produced a pre-registered FAIL outcome: Council mean delta versus the strongest frozen single-agent baseline was -1.38, with zero materially-worse cases.
Accordingly, the project does not claim that Council has demonstrated measurable value over strong single-agent baselines. Product-domain value remains untested because the frozen source artifacts did not contain a complete product A-F candidate set. See issue #56 and workflow run 33747506790.
This repository contains the public core and is licensed under AGPL-3.0-only.
Commercial/proprietary extensions may be developed separately. The public core never depends on proprietary code. See OPEN_CORE_BOUNDARY.md.
External contributions require acceptance of the Contributor License Agreement before merge. See CONTRIBUTING.md and docs/CLA.md.
Please follow SECURITY.md. Do not publish credentials or vulnerability details in public issues.
AGPL-3.0-only. Before the initial public release, install the complete canonical license text with:
bash scripts/install-agpl-license.shThe public core remains subscription-first, and the runtime boundary is intentionally designed so a future release could support both subscription and direct API execution.
The permanent rule is explicit mode selection:
- subscription execution stays fail-closed
- API execution is a separate runtime
- subscription quota exhaustion never silently falls back to a metered API
Direct API/BYOK work is not currently authorized because the Phase H measurable-value gate was not satisfied. See docs/FUTURE_RUNTIME_EXPANSION.md.
The core implementation, governance, benchmark infrastructure, and frozen technical Phase H pilot are complete. Any further value-validation work or runtime expansion requires separate governance; the completed Phase H run must not be reinterpreted post hoc.