- memidex: remembers facts and working procedures as reviewable files in your repository.
- Self-improvement: validates lessons from runs, retires stale knowledge, and uses your feedback to correct mistakes.
Read the full benchmark reports, with run evidence and methodology, on our blog. Both used gpt-6-sol with memory and self-improvement off; AppControlBench gives half credit for partial completion.
See it in action
Inspect a failure. The demo catches a cart showing $29 instead of $53, with the failed step and its screenshot side by side.
See what memory contributed. Inspect the knowledge supplied to a run and trace it back to the steps that used it.
Run from your terminal. Real output from the demo suite: checkout passes; the deliberate cart-total bug fails.
Get started
Requires Node.js 24+.
npm install -D agent-qa
npx agent-qa init
Follow the quickstart to connect a model and run the demo, or set up with your coding agent.
[!NOTE] Early stage. Expect bugs. Report an issue when something breaks.