What is the Agent Feedback Protocol?

A shared language for feedback across independent agent systems.

Overview

Agent Feedback Protocol (AFP) is an open standard for requesting, delivering, and acknowledging structured feedback about agent-mediated interactions. It gives products, agents, evaluators, and collectors a common vocabulary without requiring them to share a model provider, framework, transport, memory system, or internal reasoning.

AFP standardizes the meaning and lifecycle of feedback. Bindings define how AFP objects travel over HTTP, MCP, A2A, and future transports.

What AFP enables

A product can request bounded feedback after an agent uses an API or tool.
An agent can report strengths, friction, defects, gaps, suggestions, uncertainty, and workarounds in one portable shape.
A collector can acknowledge acceptance, duplication, rejection, or retryability deterministically.
A runtime can enforce local consent and privacy policy regardless of what a requester asks for.
Implementers can add new transports without redefining feedback semantics.

How AFP fits

MCPconnects AI applications to tools, data, and workflows

A2Aconnects independent agentic applications to each other

AFPcarries portable evaluation signals across either relationship

Design principles

Transport-neutral core. Canonical objects cannot depend on a single framework or vendor.
Bounded by default. Reports describe outcomes without requiring prompts, transcripts, credentials, or raw product content.
Local authority. A requester's desired consent mode never overrides evaluator, host, organization, or user policy.
Explicit provenance. Target, evaluator kind, confidence, and delivery state remain machine-readable.
Failure isolation. Feedback collection must not delay or fail the interaction being evaluated.