Read as deep as you need
The docs are written in layers. Start at the top: each layer assumes only the ones above it, and every page opens with its one-sentence version, so you can stop the moment you have what you came for.
- 01
The idea
Two minutes. What Evolve is, in plain words.
- 02
A change run, start to finish
One real example, step by step.
- A change run, start to finish 4 min One goal ("make checkout fast without breaking prices") got three attempts. The fastest one was wrong and got thrown out. The right one was tried on 10%, then 50% of real traffic, and then given to everyone.
- When it goes wrong 2 min A change passed every test but broke checkout for one kind of coupon. Trying it on a small share of real traffic first spotted the errors while only a few customers were on it, and Evolve put the old version back by itself.
- 03
The pieces
Each part of a run, on its own.
- Goals and the rules of the game 3 min A run starts from a goal you can measure, plus the things that must stay true. It's judged by rules that live in your repository and are locked for the whole run.
- Agents and their sandboxes 2 min Each attempt is written by a coding agent inside its own sealed workspace. Evolve records the agent's own report as untrusted, and decides only from what it measures itself.
- Checks and evidence 9 min Evolve builds each attempt and runs your checks against that exact version, in a clean, sealed-off space with no internet. Then it signs each result, so it can't be reused for anything else.
- How the winner is chosen 3 min A fixed rule from your policy (your project's rules file) picks the winner: must-haves first, then a ranking, or a reviewer agent's choice if one is set up. The decision can be replayed later and gives exactly the same answer.
- Rolling out safely 3 min The chosen version gets real traffic in small steps, compared with the current version over the same minutes. It goes live for everyone only if every step stays within the limits. If not, the old version comes back by itself.
- Experiments 2 min After a new version has been tried on a small share of real traffic, it and the current version can split real users for an A/B test (some people get version A, some get B, and you compare) on a number your business cares about, like how many checkouts become orders. The new one goes live for everyone only if it's measurably better.
- The record 4 min Every step of every run (everything Evolve does for one goal, from the attempts to going live or back) is written to a permanent record that's only ever added to. So you can always ask what's running and why, and get an answer that doesn't depend on any AI model. It's kept in Git too (the tool that keeps your code's history), so it moves between machines with your code.
- Earned autonomy 4 min Each coding agent earns, project by project, the right to put changes live without your approval, through a streak of changes that went live and stayed healthy. It loses that right the moment one of its changes causes trouble for real users, or you say you regret one.
- Where it runs 2 min Evolve works through whatever you already use to put software live, and checks for itself what's actually running. This page names the systems it supports and what each one is.
- 04
Try it
On your machine, then on your code.
- Run the demo 3 min Two commands start Evolve on your machine and take a whole change from goal to going live. A third makes one go wrong on purpose, so you can watch Evolve put the old version back. No AI key needed.
- Your first project 5 min Run evolve init, Evolve's setup command, in your project. It works out how to build it, test it and tell when it's healthy, shows you the plan line by line, and sets it up when you say yes. Then you give a coding agent a goal.
- A new project, from nothing 7 min Say what you want to build. An agent suggests a list of goals to build in order, each with its proof (what a person does in the app, step by step, or example web requests and the answers a service must give), and agents build them one at a time once you approve. Or write the example requests yourself, and they become the service's contract: the answers it must always get right.
- 05
Reference
Every setting and command.
- The policy file 14 min evolve.yaml is your project's rules file, and agents can't change it. It says what an agent may change, what counts as proof, how to pick between attempts, and how to put the winner live.
- The operator config 13 min The operator config belongs to whoever runs Evolve. It says which agents there are, where each environment is, who's allowed to call Evolve, and how much agents are trusted.
- The command line 12 min Everything Evolve does can be started, followed and questioned from the evolve command in a terminal. Evolve's web pages show the same record.
- Glossary 3 min The words Evolve uses, each explained in a sentence or two.