Expose feedback endpoints so agents can report on context quality and task outcomes

Give agents a channel to report on what worked — context relevance, task success rate, time-to-useful-context. Without feedback endpoints, you have no automated mechanism to know whether your AX is improving or degrading.

Why

The canonical AX framework (agentexperience.ax) includes "Provide Feedback" as a first-class agent-facing principle: "Agents send feedback to digital services so digital services can improve the quality of their offering and their AX. Without this information, digital services have no automated feedback mechanism to ensure the service is performing consistently."

This creates a service-side obligation: you can only receive feedback if you expose a way to receive it. Because agents are non-deterministic (the same prompt can produce different results), qualitative and quantitative feedback from agent sessions is the only reliable signal for whether your context files, tool descriptions, and API responses are actually working. Biilmann documents Netlify using agent-loop feedback to validate Dev Hub guides — if a code gen agent cannot single-shot a feature from the markdown, the guide fails.

Note: This standard is still emerging. The canonical site acknowledges it will leverage "new protocols and existing protocols/patterns (e.g. Open Telemetry)." Treat this principle as directional guidance — implement what you can with OpenTelemetry today and adapt as standards mature.

Do

Expose a feedback endpoint or accept OpenTelemetry spans from agent clients. Collect at minimum: context hit/miss rate, task success/failure, time-to-first-useful-response, and which context files or tools were used. Use this data to improve llms.txt, tool descriptions, and API responses.

Don't

Assume your AX is working because human-facing metrics look fine. Don't wait for a formal standard before collecting any feedback — instrument what you can now.

Artifact

Minimal feedback payload (OpenTelemetry-compatible attributes):

{
  "trace_id": "abc123",
  "agent": "claude-code/1.4",
  "session_id": "sess_xyz",
  "task": "deploy_site",
  "outcome": "success",
  "context_sources_used": ["llms.txt", "netlify-mcp"],
  "context_hit_rate": 0.87,
  "time_to_first_tool_call_ms": 1240,
  "tool_calls": 3,
  "tokens_consumed": 4200,
  "feedback": "llms.txt missing blob storage docs — agent fell back to web search"
}

Receive via:

• OpenTelemetry OTLP endpoint (POST /v1/traces)
• A dedicated feedback route (POST /api/ax-feedback)
• Existing analytics pipeline with agent-tagged events