Evolver

A self-improving, multi-backend agent swarm with hard cost caps — an evolutionary propose → build → verify → keep-winner loop over real objectives.

Python 3.12 · uv · SQLite · Typer + Rich · PyTorch / ROCm — private for now

Evolver is an autonomous agent swarm that runs an evolutionary loop over an objective: agents propose candidate programs, build them, verify them in a sandbox, and only the winners survive into the next generation. The first objective is GPU kernel and inference speedup, with KernelBench-style tasks — in early runs, a bundled softmax task hit a ~960× speedup over its reference implementation on an AMD RX 7900 XTX (via ROCm under WSL2).

Two design decisions do most of the work:

  • A hard cost governor. Every backend has a per-backend USD cap, a race-safe kill switch, and DB-backed rate limiting shared across all workers — the swarm can run overnight without ever silently draining a premium API quota. This was the gap in every off-the-shelf option and the reason the project exists.
  • A two-tier, game-proof verifier. Candidates are scored on public seeds, then re-verified on private held-out inputs with paired in-process timing — because agents will reward-hack a naive benchmark, and several real verifier exploits were found and regression-locked along the way.

Evolutionary loop of propose, build, verify on public seeds, re-verify on private inputs, keep winner, inside a cost-governor ring

The evolutionary loop with its two game-proof gates: private re-verification and a hard cost governor.

The swarm is genuinely heterogeneous: ~14 LLM providers behind one adapter, per-role routing (proposer / builder / verifier / auditor drawn from distinct backends), personas for division of labor, and a shared playbook that transfers accepted wins between agents. Coordination lives in a single SQLite file; a zero-dependency web dashboard watches it live.