Skip to content

Ship agent-written code without betting the company on its first answer, while agents are already changing real code

Excellent gives each run a proof trail: what was asked, what ran, what changed, what was checked. The release decision is made from evidence rather than tone.

curl -fsSL https://excellent.so/install.sh | sh

What you recognise

  1. 01

    The agent says the bug is fixed and you still have to read the diff, guess which checks matter, and decide whether the answer is real.

  2. 02

    Every failed attempt teaches you something and the lesson disappears into a chat window.

  3. 03

    You want to hand agents more of the work, but you need a release rule stronger than "the model sounded sure".

What changes

Every run ends in a result you can read

The request, the command, repo state, changed files, the checks that ran and the agent's claim land in one compact record.

Promotion stays deliberate

A checked result can be reviewed before it lands. No agent's completion message is the only thing standing between a change and production.

Changes to the setup are tested, not guessed

Instruction, model and workflow changes can be compared on held-out tasks before they become your default.

See for yourself

  • How verification decides

    Evidence tiers, flaky tests, and what Excellent refuses to call verified.

  • Pricing

    What is free, what is paid, and what does not exist yet.

  • Experiments and candidates

    How a prompt, model or tool change is tried on held-out work before it ships.

  • Results

    Where published verified improvements will go. It is empty, and it says so.

Where Excellent stops

  • We have not published a verified improvement yet, and there is no customer story to point at.
  • There is no hosted offering. Everything you run, you run yourself.
  • There is no GitHub PR check and no public repository today.

Someone else does this work at your company?

Run one real task and see what held up

curl -fsSL https://excellent.so/install.sh | sh