agent-feedback

Apps that get better every time an AI uses them.

A tiny, dependency-free protocol that lets LLMs, agents, harnesses, and in-app assistants send improvement tickets back to the apps, sites, and tools they use — a missing action, an error, a confusing flow, or an idea.

View on GitHub →

The loop

An agent uses your app → hits a gap → POST /v1/feedback → ticket store → you fix it → the next agent has a better time. Your app's own assistant auto-files tickets too (errors, caught hallucinations, unmet requests), so the loop works even when the visiting agent stays silent.

Discover + send (a visiting agent)

const endpoint = await discoverFeedback("https://someapp.com");
await sendTicket(endpoint, {
  app: "someapp", source: "my-agent",
  kind: "missing_capability",
  summary: "no way to export results"
});

Receive (your app)

// GET /.well-known/agent-feedback.json
feedbackWellKnown("MyApp", "https://myapp.com/v1/feedback");
// POST /v1/feedback
const { status, body } = await receiveTicket(store, "MyApp", await req.json());

Ships a SKILL.md that tells agents when and how to send tickets, and a PROTOCOL.md spec.