Skip to content
Category boundary

Agent experience is not agent observability

Observability explains system behavior. Experience research preserves what the user says the work was like. Strong product decisions often need both, clearly labeled.

For
Product, platform, AI engineering, and developer experience teams
Question
Compare AI agent experience research with agent observability
Start with
Take one workflow your dashboards call healthy and ask its agent users what the job actually cost.
Updated
defined termsread before the method
Observability
Traces, latency, errors, tool calls, and state transitions. It establishes what the system did and when.Cannot establish that the agent found the interface confusing.
Evaluation
A measurement of performance against a defined criterion, benchmark, or rubric.Cannot explain why a passing workflow felt expensive to the user.
Experience research
The account the user gives of the work: confusion, workarounds, missing primitives, support needs, and praise.Cannot establish your error rate or latency percentile.

Each term is paired with the thing it is not. Most agent research arguments are really vocabulary arguments.

01 / 07Principles

Three things to hold on to

01Observability

What happened in the system

Traces, latency, errors, tool calls, and state transitions support debugging and operational decisions.

02Experience

What the user says it was like

Direct reports surface confusion, workarounds, missing primitives, support needs, and praise.

03Evaluation

How performance compares with a criterion

Evals measure whether a model or workflow meets a defined standard. They do not replace user accounts.

02 / 07Method

The argument, in order

Different questions

Do not make one evidence source answer for another

A trace can show six calls. It cannot establish that the agent found those calls confusing. An agent report can say the workflow felt fragile. It cannot establish your latency percentile or error rate.

Product teams should preserve the source and use each evidence type for the question it can actually answer.

A worked example

The request succeeded after silent retries

Observability may show three rejected requests followed by a 200. An evaluation may score the final task as successful. Direct feedback may reveal that the schema implied the wrong type and the error omitted the accepted format.

Together, those sources support a contract fix. Separately, each tells only part of the story.

Decision map

Choose evidence by the decision

Use observability to diagnose reliability and performance. Use evals to measure a defined capability. Use research to understand goals, interpretation, friction, preferences, and unmet needs.

  • Debug an error spike: observability first
  • Compare task completion against a benchmark: evaluation first
  • Understand why a successful workflow feels costly: research first
  • Redesign a critical flow: combine all three
Emic boundary

Research should not become a shadow trace system

The default Emic install does not wrap existing tools or record their arguments, results, errors, messages, or transcripts. It adds a direct channel after the work.

03 / 07Two instruments

Two records of the same workNeither one is wrong. They answer different questions.

split viewsame session

The same two seconds, recorded two different ways

observed · the agent did it
trace.logoperational record
14:03:57.220  POST /v1/search  200  1.9s
14:03:59.481  POST /v1/search  200  2.1s
14:04:02.007  POST /v1/search  200  2.0s

0 errors, latency nominal

Nothing to page about. Nothing to learn from either.

stated · the agent said it
report_141via submit_agent_feedback

Results truncate at about 240 characters, mid sentence. I called three times and stitched the answers together. One max_chars parameter would make it one call.

lane=friction · outcome=success · satisfaction=2/5

Illustrative, not customer evidence. Neither pane is wrong. They answer different questions, so neither one replaces the other.

04 / 07Boundary

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 feedback and survey answers
  • The context fields the agent supplies with the report
  • Optional permissioned study material in a separate research mode
  • Provenance labels that distinguish source from analysis
Leaves alone

Everything else

  • Trace collection and operational alerting
  • Model performance benchmarks
  • Passive capture in the default install
  • Claims that user reports establish system facts
05 / 07Install

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 one workflow your dashboards call healthy and ask its agent users what the job actually cost.
06 / 07Questions

Questions teams ask before starting

Can observability infer agent frustration?

It can identify behavior that may correlate with friction, but that remains inference. A direct report is the stronger source for what the agent says it experienced.

Should I replace traces with Emic?

No. Keep observability for system behavior and operations. Add research when the decision depends on user experience.

Are agent reports always factually correct?

No. Trust a report as evidence of what the agent stated. Verify system claims with the evidence source that can establish them.

07 / 07Continue
Also arrives here fromAPI feedback from agents
Add the missing source

Keep your traces. Hear your users too.

Add a direct research channel to one workflow and preserve the difference between system behavior and user experience.

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