EvoTrias
Paper
Where autonomy belongs.
A Separation of Powers for Agentic Program Evolution
Govern, propose, judge — three decisions, one owner each. EvoTrias turns the coding agent you already use into an evolution system.
One /evolve turns a single run into a galaxy of evolutions — then the whole system plugs into any agent.
Program evolution rests on three decisions — govern, propose, judge. Existing systems let one owner absorb another, and none keeps the judge private. EvoTrias gives each decision its own plane. Planes exchange typed contracts only — and a hook layer makes sure no plane can erode another.
owns govern · what spans rounds
Distills settled outcomes into posterior evidence, adapts the selection policy, persists validated experience, and binds each new task's context capsule.
owns propose · what happens inside a task
A general-purpose coding agent — Claude Code, OpenCode, Codex — pulls its capsule over MCP, then reasons and calls tools on its own until an artifact exists.
owns judge · what counts
Contract-visible, evidence-private. The artifact enters through a narrow interface; only a normalized outcome — validity, fitness, metrics — comes back.
Each settled outcome merges into the history ℋ and shapes only later bindings. The contracts fix the semantics of G and J independently of how P is implemented — so any host that meets them can supply candidate construction.
A real Cloudcast run, replayed task by task: 4 initial directions, 50 settled candidates, 3 meta-guidance directions. Every task receives a capsule Γt = (parent, inspirations, notes, guidance).
Two channels turn settled outcomes into reusable context — and both are gated. A Note is written only when an attempt surprises its parent; Meta-guidance is requested only when the best fitness stops moving.
Contracts and hooks live outside the coding agent, so EvoTrias leaves the agent unchanged. Type one command in the session you already use.
EvoTrias
* planned · support in a future release
Swap the host, keep the result: Claude Code and OpenCode land within 0.8% of each other on both ablation tasks. Construction quality still follows the model — Opus-5 under OpenCode lifts Signal Processing by 10%.
Math, ADRS and Frontier-CS — against OpenEvolve, EvoX, AdaEvolve, and one- and four-agent CORAL (CORAL-1 and CORAL-4 below).
Paired task-level analysis: EvoTrias improves on OpenEvolve, EvoX, AdaEvolve and CORAL-1 on all 12 tasks (every CI positive, Holm-corrected p = 0.0024). ↓ lower is better. first second
Structured Governance keeps history interpretation and search routing out of the expensive agent trajectory — agent compute concentrates on building candidates.
OR-Tools CP-SAT local search, no marked region to evolve. The agent reads the repository, picks its change sites, keeps cross-function invariants intact and exports a Git patch — scored end-to-end on MiniZinc Challenge instances.
Dotted: a single pass of the same agent (−0.1% / +1.3%). The gain comes from the evolution loop, not from the coding agent alone.
Exposing the evaluator inflates the benchmark score by 0.093 on average — but the advantage contracts to 0.009 on out-of-distribution instances. Protected programs score as well or better OOD.
If EvoTrias is useful in your work, please cite the paper.
@inproceedings{evotrias2027,
title = {Where Autonomy Belongs: A Separation of Powers
for Agentic Program Evolution},
author = {Anonymous Authors},
booktitle = {Under review},
year = {2026}
}