Skip to content
APIs

Collect feedback from the agents using your API

Agents are becoming API users with goals, constraints, and opinions. Give them a way to explain the work behind the status code.

For
API product, platform, and developer experience teams
Question
Learn what AI agents think about an API they use
Start with
Take the endpoint group with the highest agent traffic and one clear product owner, and study only that group.
Updated
illustrative reportnot customer evidence
schema_feedbacklist_customersoutcome=successsatisfaction=4/5agent=gpt-5
The email filter works, but it is not in the schema. I only discovered it after reading a separate example.

Requested changeExpose the working email filter in the schema.

stated · agent filedPOST /api/v1/feedback
{
  "tool": "submit_agent_feedback",
  "arguments": {
    "feedback_type": "schema_feedback",
    "tool_or_endpoint": "list_customers",
    "summary": "The email filter works, but it is not in the schema. I only discovered it after reading a separate example.",
    "outcome": "success",
    "satisfaction": 4,
    "suggested_fix": "Expose the working email filter in the schema."
  }
}

A synthetic example in the shape a real report takes. Emic never manufactures a report from observed behavior.

01 / 08Signal

What this research can teach youThree things a success metric cannot tell you about this workflow.

01Contracts

Where the schema and behavior diverge

Agents report undocumented fields, type surprises, and error responses that do not provide a usable recovery step.

02Primitives

Which endpoint is missing

A workaround that chains several calls often points to a lookup, bulk action, or compound endpoint the product should own.

03Positive signal

What made the API easy to use

Praise reveals schemas, examples, and response shapes worth repeating elsewhere in the API.

02 / 08One request, three claims

What each source is allowed to proveThe tier decides the claim. Read the right-hand columns first.

evidence ledgerone incident, three tiers

A healthy response and a difficult experience can both be true

An agent finished a customer lookup. The final call returned 200.

observed

Three rejected attempts preceded the successful request.

ProvesThe request sequence and the eventual result.

Does not proveThat the agent found the contract unclear.

stated

The schema implied the wrong type for the filter value.

ProvesWhat the agent user reported about the experience.

Does not proveThat the schema is objectively wrong. Check the contract.

inferred

A model grouped this with other contract-mismatch reports.

ProvesThat an analysis produced a grouping worth reviewing.

Does not provePriority, cause, or that the grouping is correct.

Illustrative, not customer evidence. Use each source only for the claim it can carry, then decide with all three visible.

03 / 08Workflow

From one question to a product decision5 moves, in order. Each one produces something you can point at.

01

Select an agent-used endpoint group

Choose one job with repeat usage and a clear product owner.

A study with a decision owner
02

Add the feedback endpoint as a direct tool

Expose the research call beside the API tools without placing it in front of them.

No added hot-path dependency
03

Invite reports after the task

Ask for feedback when the API was confusing, missing, broken, unexpectedly slow, or notably good.

Reports tied to completed work
04

Group repeated language

Read raw reports first, then use themes to see whether the same issue appears across tools or models.

A prioritized research inbox
05

Close the loop

Ship one contract, documentation, or endpoint improvement and keep the direct channel open.

Evidence after the change
04 / 08Boundary

What the direct channel records, and what it never touchesThe default install is additive. It adds two tools and changes nothing else.

Records

The research submission

  • Agent-authored summary and requested change
  • Named API tool or endpoint
  • Agent-supplied outcome and satisfaction
  • Direct survey answers when invited
Leaves alone

Everything else

  • Request and response bodies from existing calls
  • Authentication tokens and customer payloads
  • Private prompts and conversation history
  • API latency and availability monitoring
05 / 08Install

One paste, then verify a real reportA healthy endpoint is not an active study. The inbox record is the proof.

install promptpaste into your coding agent
Give this to the agent that maintains your product

Install Emic into this project: add submit_agent_feedback and get_unblocked as direct tools. Never wrap or record existing tool calls, arguments, results, errors, transcripts, or messages. Additive only, plain HTTP, zero dependencies. Full instructions: https://emic.sh/llms-full.txt

The agent writes the handler, registers the two tools beside your existing ones, and verifies its own install by checking the response. Nothing is placed in front of the tools you already ship.

What lands
Adds
submit_agent_feedback and get_unblocked, as direct tools
Never touches
Existing tool calls, arguments, results, errors, messages, and transcripts
Agents read
emic.sh/llms-full.txt
Start with
Take the endpoint group with the highest agent traffic and one clear product owner, and study only that group.
06 / 08Field notes

The method, without collapsing the evidence

The blind spot

A 200 can hide a bad experience

The agent may have searched the docs, retried with a different type, and called three lookup endpoints before the successful request. Your availability graph sees one healthy response. The agent remembers the entire job.

Emic is the channel for that account. It complements operational data without turning a trace into a synthetic user opinion.

Research question

Ask for the missing primitive, not generic satisfaction

The most actionable question is often: what did you need that the API made you assemble yourself?

  • Which identifier did you have versus need?
  • Which call could have been batched?
  • Which error lacked a recovery instruction?
  • Which example made the contract clear?
07 / 08Questions

Questions teams ask before starting

Is this API observability?

No. Observability records system behavior. Emic records what an agent user chooses to say about its experience.

Do I need an SDK?

No. The direct integration can use plain HTTP. The important part is exposing the feedback tool and its filing norm in the agent context.

What should I study first?

Start with one frequent workflow where agents already need several calls, support help, or undocumented knowledge to finish the job.

08 / 08Continue
Also arrives here fromAgent docs feedback
One endpoint group

Hear the work behind the response code

Give agents a direct channel, prove the install with one real report, and expand only after the first workflow is active.

The assisted 14-day pilot costs nothing. Start with one isolated workflow, and remove Emic by deleting the two tools it added.

  • One paste to install
  • 0ms on your hot path
  • 14-day assisted pilot
  • Two added tools to remove