EvolveSource

01 The idea · 3 min read

Evolve in two minutes

Coding agents are AI tools that write changes to your code for you. If you’ve got some working for you, this page is for you. In two minutes you’ll know what Evolve is, where it fits, and whether it’s worth ten more minutes of your time.

The problem

Coding agents like Claude Code or Codex can write real changes to real software now, and fast. That’s great. It also leaves you with an awkward choice, every single time:

  • Read every line yourself. Safe, but now everything waits on you, and the agent’s speed is wasted.
  • Trust it. Fast, right up until the day the agent says “all tests pass” (tests being the small programs that check your code does what it should) and something breaks for the people using your software anyway.

Neither is much fun. Evolve is a third option.

The idea: it’s a hiring process

Think about how a careful team hires someone.

  1. They say what the job is, clearly enough to judge it. Something like “answer support tickets within a day”, which you can actually check.
  2. They see a few candidates. One applicant is a gamble. A few give you a choice.
  3. They check the work themselves. The team marks a work sample itself.
  4. They pick by a rule they agreed on beforehand.
  5. They start the new hire on probation. Small jobs first, watched closely, with a way back if it goes wrong.
  6. They keep records, so a year later anyone can see why this person got the job.

Evolve does exactly this for changes to your software. The “candidates” are changes written by coding agents. The “job” is a goal you can measure, like make checkout faster without getting any price wrong.

your agents propose changes evolve 1 · each attempt sealed off on its own 2 · checks it cannot edit 3 · one candidate chosen, by rule 4 · a slice of real traffic, then all production via your deploy tool every step written to a record you can query: evolve why
Figure Evolve sits between your agents and production. It works with your Git repository (Git keeps your code and its history), your tests and the tool you deploy with.

What that looks like day to day

You give Evolve a goal. Agents come back with a few different changes, each made in its own sealed-off workspace. Evolve checks every one with tests the agents can’t see or change, and picks the winner by a rule written in your repository (the place your project’s code and its history live).

Then comes probation. The winner gets a small share of real users, right next to the version that’s already running. If it does well, it gets more. If it doesn’t, Evolve puts the old version back by itself.

Every step is written down for good. So later you can point at anything running in production, the live version real people use, and ask: why is this here? Evolve answers with the goal, the evidence, the decision and what it saw happen.

Where it fits

  • It works with the coding agents you already use: Claude Code, Codex, opencode, or any agent that can work without a person at the keyboard.
  • It uses what you already have: Git (which keeps your code’s history), your tests, your CI (the build and tests that run on every change) and whatever you use to put new versions live.
  • You still make the calls. You decide the goal and what counts as proof. Evolve makes sure only proven changes get through, and shows you exactly what was proven.

Where to go next