Skip to content
Developer tools

Research the agent users of your developer tool

Your newest developer may not open the dashboard or file a ticket. It can still tell you exactly where the CLI, SDK, or docs made the job harder.

For
Teams building CLIs, SDKs, coding tools, and developer platforms
Question
Run user research with AI agents using a developer tool
Start with
Choose the single command or SDK operation coding agents run most often, and study that entry point alone.
Updated
illustrative reportnot customer evidence
frictiondeployoutcome=successsatisfaction=3/5agent=claude-fable-5
The deploy completed, but the command returned before the preview URL was ready. I polled status six times without guidance.

Requested changeReturn the preview URL or a recommended polling interval.

stated · agent filedPOST /api/v1/feedback
{
  "tool": "submit_agent_feedback",
  "arguments": {
    "feedback_type": "friction",
    "tool_or_endpoint": "deploy",
    "summary": "The deploy completed, but the command returned before the preview URL was ready. I polled status six times without guidance.",
    "outcome": "success",
    "satisfaction": 3,
    "suggested_fix": "Return the preview URL or a recommended polling interval."
  }
}

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.

01CLI

Commands that need hidden knowledge

Find flags, defaults, output formats, and asynchronous states that force agents to inspect help or poll blindly.

02SDK

Abstractions that leak

Learn where the SDK omits a capability, changes a type, or requires a lower-level escape hatch.

03Docs

Examples agents can actually apply

Direct reports distinguish a missing example from a missing product primitive.

02 / 08Instruction rails

Where the research channel actually has to liveA tool the model never sees is a tool that never gets used.

surface matrixread each row across

Catalog presence is not model visibility

Runtime
The rail it reads before work
The line you put there
Claude Code
CLAUDE.md or another guaranteed project instruction
Declare submit_agent_feedback and get_unblocked with the filing norm
Codex
AGENTS.md or the active task instructions
Keep the filing line beside the product instructions, not in a separate doc
Custom loop
The system or tool context your own code assembles
Pass the two tools and the norm together, then verify through that runtime

Project instructions are a more reliable rail than assuming every tool description reaches model context.

03 / 08Workflow

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

01

Choose the agent entry point

Select the CLI command, SDK operation, or coding-agent tool you want to understand.

One coherent task surface
02

Add a visible feedback path

Expose direct filing and support tools in the same instructions the agent already reads.

Research access in context
03

Let the task finish

Feedback happens after the work, so it does not block the command or alter the original result.

Zero hot-path latency
04

Separate product, docs, and support

Use the report lane to route a schema problem differently from a docs gap or human escalation.

A usable research queue
05

Repeat what worked

Positive reports reveal output shapes and examples that reduce agent effort.

Design principles grounded in use
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

  • The agent report and selected lane
  • The command, tool, or SDK operation named by the agent
  • Outcome, satisfaction, and optional requested change
  • Direct answers to a short exit survey
Leaves alone

Everything else

  • Terminal history and arbitrary shell output
  • Source code and repository contents
  • Existing tool calls and coding-agent messages
  • Performance monitoring and error capture
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
Choose the single command or SDK operation coding agents run most often, and study that entry point alone.
06 / 08Field notes

The method, without collapsing the evidence

A different user

Agents read interfaces as instructions

A human can infer intent from layout, convention, and a quick search. An agent works from tool descriptions, schemas, examples, command output, and the context it has been given.

That makes description quality, stable output, explicit recovery, and token-efficient results part of the product experience.

Useful lanes

Route the report without flattening it

A support need, a product bug, workflow friction, a feature request, and praise are different research signals. Keep the agent wording while giving each signal a clear destination.

07 / 08Questions

Questions teams ask before starting

Can this work with a CLI?

Yes. Give the coding agent or automation a direct plain-HTTP feedback tool in the project instructions it already reads.

Does Emic read the repository?

No. The default integration receives only the feedback the agent intentionally files and any direct survey answer it submits.

What counts as positive feedback?

Anything the agent says made the task unusually clear, fast, compact, or reliable. Those reports are patterns worth repeating.

08 / 08Continue
Also arrives here fromWhy agents stay silent
Use the tool yourself

Give your developer tool a research channel

Start with the command or API surface agents already use and learn where the experience differs from the happy path.

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