Redteam pits two AI models against each other in a TDD code-review pipeline
Redteam is an open-source adversarial agent-pair harness where one AI model writes code through a plan→implement→review pipeline while a second model reviews it adversarially, gated by human checkpoints before merge.
Score breakdown
Redteam introduces a human-gated, dual-model review structure that directly counters the single-model blind spot of an AI both writing and approving its own code.
- 01One AI model writes code through a plan → implement → review pipeline; a second model reviews it adversarially.
- 02The system is designed to prevent automatic self-agreement that occurs when a single model reviews its own output.
- 03Human checkpoints gate the output — the final result is a draft PR a human reviews before merge.
Redteam is an adversarial agent-pair harness for AI-assisted code shipping, published by AscendyProject under the Apache-2.0 license. The core design pairs two independent AI models: one drives a task through a fixed pipeline of plan, implement, and review phases, while a second model reviews the work adversarially. The deliberate collision of two independent model perspectives is described as the point of the system — it exists specifically to prevent the automatic self-agreement that would occur if a single model both wrote and reviewed its own output. The pipeline persists `state.json` after each phase, making runs fully resumable and enabling retries when a review returns `CHANGES_REQUESTED`. The final output is a draft PR that a human reviews before merge, preserving a human checkpoint before any code lands.
The repository is described as early-stage, with APIs and layout still subject to change.
The project also offers a single-model TDD mode that front-loads `write_test → verify_test` for teams not running the full dual-model setup. Redteam was originally built as an internal harness for one project and has driven real, merged pull requests before being extracted into this standalone repo. The repository is described as early-stage, with APIs and layout still subject to change. A quick-install path is available for Claude Code users via two `/plugin` commands, with a vendor-into-any-repo option documented for other environments.
Key facts
- 01One AI model writes code through a plan → implement → review pipeline; a second model reviews it adversarially.
- 02The system is designed to prevent automatic self-agreement that occurs when a single model reviews its own output.
- 03Human checkpoints gate the output — the final result is a draft PR a human reviews before merge.
- 04`state.json` is persisted after each phase, making runs fully resumable with retry support on CHANGES_REQUESTED.
- 05A single-model TDD mode (write_test → verify_test) is also available.
- 06The project was originally an internal harness that drove real, merged pull requests before being extracted into this standalone repo.
- 07Licensed Apache-2.0; described as early-stage with APIs and layout still subject to change.
Topics
Summary and scoring are generated automatically from the original article. We always link back to the publisher and never republish images or paywalled content. Last processed Jun 19, 2026 · 10:25 UTC. How this works →