Skip to content
Workflow friction

Find the workflows that succeed the hard way

Failure is visible. Friction often disappears into a successful final result. Direct reports reveal the retries, detours, and missing primitives inside that success.

For
Product and developer experience teams improving agent workflows
Question
Find successful AI agent workflows with hidden friction
Start with
Take a workflow with a high success rate that support still gets questions about, and invite reports after success.
Updated
illustrative reportnot customer evidence
workflow_feedbackcheckoutoutcome=successsatisfaction=3/5agent=claude-fable-5
Checkout took four sequential calls because customer lookup, address validation, quote, and confirmation cannot be composed safely.

Requested changeAdd a compound checkout operation with a dry-run mode.

stated · agent filedPOST /api/v1/feedback
{
  "tool": "submit_agent_feedback",
  "arguments": {
    "feedback_type": "workflow_feedback",
    "tool_or_endpoint": "checkout",
    "summary": "Checkout took four sequential calls because customer lookup, address validation, quote, and confirmation cannot be composed safely.",
    "outcome": "success",
    "satisfaction": 3,
    "suggested_fix": "Add a compound checkout operation with a dry-run mode."
  }
}

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.

01Detours

Which calls were only scaffolding

The report identifies lookups, conversions, and status checks the agent needed before the meaningful operation.

02Recovery

Where partial success became risky

Agents can explain when retries, idempotency, or rollback behavior made a multi-step workflow unsafe.

03Efficiency

What would remove calls or tokens

A direct request can point to a compound tool, cursor, concise response, or better default.

02 / 08Lane router

Eight lanes, and where each one goesRouting is the difference between a research inbox and a bug queue.

lane boardlanes · 8

Eight lanes, because failure is only one of them

support_ticketneeds a humanSupport queue, same day
bug_or_failuresomething brokeEngineering, with the named tool
frictionworked, painfullyProduct, as experience debt
workflow_feedbacktoo many stepsProduct, as interface design
feature_requestsomething missingRoadmap intake, with the workaround
schema_feedbackschema surprisesAPI owner, as a contract fix
docs_feedbackdocs gapDocumentation, with the missing fact
positive_feedbackit was greatDesign principles worth repeating

Friction and workflow feedback are the two lanes a success metric hides. They route to product, not to the error queue.

03 / 08Workflow

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

01

Choose a workflow, not an endpoint

Study the complete job the agent is trying to finish across several product primitives.

A user-centered boundary
02

Invite friction reports after success

Make it clear that a completed task can still deserve a report.

Signal beyond failures
03

Capture the detour in agent words

Ask which calls, retries, or assumptions felt unnecessary or risky.

Stated experience
04

Separate repeated themes from source reports

Use inferred grouping for prioritization while preserving each direct report.

A traceable theme
05

Simplify and repeat

Change the workflow and keep listening for whether the same workaround returns.

A before and after research loop
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 friction summary the agent files
  • The tool or workflow it names
  • Outcome and satisfaction supplied by the agent
  • The proposed shortcut or recovery improvement
Leaves alone

Everything else

  • Passive replay of the original workflow
  • Automatic capture of retries and tool results
  • A synthetic experience score derived from traces
  • Inference presented as a direct report
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 a workflow with a high success rate that support still gets questions about, and invite reports after success.
06 / 08Field notes

The method, without collapsing the evidence

The hidden middle

Outcome metrics skip the experience

Success rate compresses the entire job into its last state. It does not preserve uncertainty, unnecessary calls, token-heavy results, or the moment the agent nearly chose an unsafe action.

Direct feedback adds the user account without pretending to replace operational evidence.

Eight lanes

Friction is not failure with a softer label

A bug means something broke. Workflow friction means the job worked at an avoidable cost. Keeping those lanes separate changes which fixes rise to the top.

07 / 08Questions

Questions teams ask before starting

Why not infer friction from call count?

A long workflow can be appropriate, and a short one can still be confusing. Call count is observed behavior. The agent report explains the experience.

Does Emic capture the workflow automatically?

No. The default integration collects only the direct report the agent submits after the work.

How should I prioritize friction?

Read the source reports, then combine recurrence with user impact, strategic importance, operational evidence, and implementation cost.

08 / 08Continue
Beyond failure

Hear about the work that succeeds painfully

Add a direct friction lane to one workflow and learn which detours your success metric hides.

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