Huawei's EvoTrias: separating search, code and scoring cuts agent token cost 64%, lifts OR-Tools 13.2%
rohanpaul_ai · x · 2026-11-17
A Huawei Taylor Lab paper introduces EvoTrias, a 'separation of powers' architecture for AlphaEvolve-style program evolution: a controller steers the search, a coding agent like Claude Code proposes code, and a protected evaluator the agent cannot read scores it. On 7 math tasks it led 6 and cut token cost per run by 64% vs the 4-agent CORAL system. On OR-Tools, a 680K-line C++ repo, it improved solver quality 13.2% across 300 instances, while 1-shot runs of the same agent averaged -0.1%. Takeaway for code-optimizing agent loops: hide the scorer from the agent and keep search planning out of the coding agent. Code is open-sourced.
More from coding & agent
- YC amplifies viral take: 99.999% of 2023 'software development' is automated, yet SWE openings are up — ycombinator · 2026-11-17
- Agent orchestration: weak model at front desk, strong models in worker threads — dotey · 2026-11-17
- Huawei's EvoTrias: separating search, proposal, and judgment powers sets new SOTA — rohanpaul_ai · 2026-11-17
- agent-browser: open-source browser automation so agents can verify fixes and file PRs — evilrabbit_ · 2026-11-17
- OpenRouter compares the context-token cost of MCP, CLI and Skills — rseroter · 2026-11-17
- PSN: storing agent skills as executable programs in an evolvable network — burkov · 2026-11-17