EvolveSource

05 Reference · 3 min read

Glossary

If you’ve hit a word you don’t know, this page is for you. It’s probably here. Each one gets a plain sentence or two, and a link if there’s a whole page about it.

Agent (or coding agent): an AI tool, like Claude Code or Codex, that writes changes to your code.

Attempt: one agent’s try at a goal, made in its own sandbox. A run usually makes a few.

Autonomous: an agent whose chosen changes ship without a person approving them. It’s earned, and it can be lost. See earned autonomy.

Branch: a named line of work in a Git repository, like main. Your main branch is the version of the code everyone works from.

Canary: the new version running beside the current one on a small share of real traffic, and being watched. See rollout.

Candidate: an attempt that produced a change worth checking.

Check: one trusted test of a candidate. It’s defined in the policy and run with your own protected copies of the files. See checks.

Check escape: a candidate that passed every check but then got caught in production, or regretted. It means the checks missed something.

CI (continuous integration): the build and tests that run on every change to your code, wherever you already run them.

Commit: a saved snapshot of your code in Git.

Confidence: how well Evolve is living up to what you expect of it, which is that only proven changes ship by themselves. It’s measured by indicators against objectives.

Decision: the recorded choice of which candidate, if any, goes forward, with the reasons. It can be replayed from the evidence.

Diff: the line-by-line list of what a change did.

Digest: the fingerprint of a built program. Evolve tells versions apart by digest, so “what’s running?” has an exact answer.

Evidence: the signed result of a check: what ran, against what, and what happened.

Experiment: an A/B trial of the candidate against the current version on real users, measuring something the business cares about. See experiments.

Goal (or objective): what a run is for, written as a target you can measure plus the things that must stay true. See goals.

Indicator: something Evolve measures about itself, like the check escape rate.

Known-good: the version Evolve goes back to if anything goes wrong. There’s always one before a change is tried.

Operator config: the settings of Evolve itself, owned by whoever runs it. See operator config.

p95: how long the slowest 5% of requests take to answer. A common way to measure speed.

Policy: evolve.yaml, the repository’s rules for what agents may change and what counts as proof. See the policy file.

Production: the live version of your software, the one real people use.

Promotion: the candidate taking all the traffic and becoming the new known-good version.

Protected path: a file or directory no attempt may change, like your tests and the policy itself.

Real traffic: requests from the real people using your software.

Record (or ledger): the permanent account of everything Evolve did. It’s only ever added to. See the record.

Regret: a person reporting that a promoted change turned out to be harmful after all.

Repository: the place your project’s code and its history are kept, in Git.

Rollback: sending traffic back to the known-good version.

Run: one goal, from the attempts through to promotion or rollback.

Sandbox: a sealed-off space a program runs in, cut off from the rest of the machine.

Supervised: an agent whose chosen changes wait for a person’s approval.

Threshold: the level of an indicator at which Evolve steps in, by putting agents back under supervision.

Toolchain: the tools a project is built and tested with: Go, Node, Python or Rust. The policy names it, and checks and agents are given it.