Chimera tests a cost-aware way to build AI agents
Chimera is an agent released under Apache-2.0. For hard tasks, several LLMs answer the same prompt, a judge model compares where they agree or conflict and what they missed, and a final model combines that review into one answer. Easy tasks and tool-use steps stay on a single model, so a cost-and-speed router only spends extra money on multiple models when it is expected to help. The agent loop plans work, acts, checks the result, and rolls back work that fails .
Verified execution results are treated as the , so a manager component is not allowed to throw away work that has already been proven correct. Its stores and searches past facts and sessions with SQLite and FTS. It also includes rules for allow, warn, block, and review decisions, plus a static check for cases where the agent tries to modify itself. Chimera can connect to tools through MCP and OpenAPI imports, and it can split work across isolated subagents that each pass their own gates.
LiteLLM lets it work with model endpoints, including . The project is still alpha: it has 469 tests and passes strict code checks, but it has not yet been proven in production use.
Key points
- Hard steps use several LLMs, while easy steps stay on one model.
- A judge agreement, conflicts, missing coverage, and weak spots before the final answer is written.
- A cost-and-speed router decides when the extra are worth paying for.
- The agent verifies work and can roll back failed actions.
- It has 469 tests, but no production track record yet.