Make your agents prove their work.

The bottleneck moved from generation to validation.

AI agents now write code faster than any team can review it. An agent that writes its own tests and then grades itself produces output that agrees with itself — fluent, green-looking, and wrong often enough to break production.

And the naive fix — send everything to a frontier model, run it many times — makes the bill grow faster than the unit price falls. Coevolve is built for the world after “make it generate.”

of enterprise technology leaders report an increase in production issues linked to AI-generated code.

cost cut by routing between a strong and weak model, without compromising quality — the lever Coevolve automates against a verification bar.

Most agent pilots stall on the same three things — evaluation, reliability, and governance — which Coevolve treats as runtime primitives.

It verifies correctness — it doesn’t just orchestrate

It governs cost across a pool of models

It learns from what actually verified

119 shipped capabilities, seven pillars

git clone https://github.com/SourceShift/coevolve.git
cd coevolve
make install
# new terminal, then confirm .venv
coevolve version

# install OS prereqs with winget first
# (Python 3.11, Git, jq, yq, SQLite)
py -3 .\scripts\full_install.py
coevolve version

# remember the checkout, then make a project
COEVOLVE_SOURCE="$PWD"
mkdir -p ~/coevolve-demo && cd ~/coevolve-demo
git init && coevolve init

# a kickoff = goal + scope + artifact + proof
cp "$COEVOLVE_SOURCE/examples/01-hello-world/kickoff.md" ./kickoff.md

# run locally, without provider calls
COEVOLVE_DRY_RUN=1 coevolve run code-fix ./kickoff.md
coevolve validate

State the target repository, allowed files, intended artifact, and the command or rule that proves success.

It checks the lifecycle and artifact paths without model calls — it does not prove the eventual change is correct.

Prefer an existing test, typecheck, schema, fixture, or observable acceptance criterion over an LLM-only score.

Heterogeneous review is useful for discovery and diagnosis; it does not replace deterministic verification.

Traces are retained and learned from, but automatic promotion is restricted to classes with measurable external evidence.

Prefer a real check over a confident opinion. Where there isn’t one, surface the uncertainty.

{{ activeRecipe.desc }}

recipes live in recipes/ — task class + workflow + artifact contract + prompts + verifiers

Truthful dispatch telemetry, an error and finish-reason taxonomy, heartbeat/failure handling, capability-aware routing, cost accuracy, and operator intervention policies.

Build a library of recovery behaviors from failure analysis, then learn a policy that chooses among a cheap tool call, more planning, a stronger model, or a user interruption — with verified progress as the signal, balanced against compute, latency, and the user’s interruption budget.

Coevolve is Apache-2.0 and early. Use a dedicated worktree for framework changes, keep a verifier with every behavior claim, and run the focused checks for the surface you change. The workflow and quality gates live in AGENTS.md; direction in GOVERNANCE.md.