# How To Audit Your Agent Experience

Source reference:

- [Getting Started with Agent Experience](https://agentexperience.ax/guides/getting-started-guide/)

## What The Source Is Saying

- Start with assessment before changing the system.
- Review current tools, processes, performance, and pain points.
- Plan improvements with clear objectives and success measures.
- Implement changes, test with a smaller group or workflow, gather feedback, and keep iterating.

## What Marco Takes From It

The guide is broad, but the sequence is useful.

Assess first.

Then improve.

Designers need the same move with agents. Do not start by installing more tools. Do not start by writing a giant prompt. Start by looking at one workflow and asking where the agent gets lost.

That is usually where the Harness begins.

## Pick One Agent Workflow

Choose one repeated task.

Not the whole design process.

One loop.

Examples:

- critique a mobile screen
- turn a rough brief into an HTML mockup
- check a UI against the design system
- summarize research notes
- verify source trust before writing product copy
- audit loading states
- prepare a handoff note

The smaller the loop, the easier it is to see the failure.

## Assess What The Agent Cannot Understand

Run the workflow once and watch where the agent misses.

Ask:

1. Did it know the product goal?
2. Did it know the user context?
3. Did it know which source to trust?
4. Did it know the design rules?
5. Did it know what output shape you needed?
6. Did it know what required human judgment?
7. Did it make the same mistake you have corrected before?

This is not about blaming the model.

It is about finding the missing harness surface.

## Plan The Missing Context

Each failure points to a fix.

| Failure | Harness fix |
| --- | --- |
| Agent invents product facts | add source-trust rules |
| Agent ignores visual tone | improve `DESIGN.md` |
| Agent changes too much | add constraints and scope rules |
| Agent gives vague critique | create a critique skill |
| Agent cannot finish safely | add escalation rules |
| Agent repeats old mistake | update failure log and checklist |

The fix should live outside the chat.

If the lesson stays only in the conversation, it will disappear.

## Test With A Real Task

Do not test the Harness on fake work.

Use a real task with real constraints.

That is why Planark and the Lab matter. Real work exposes fake confidence. It forces source trust. It shows when design polish is hiding a product mistake. It shows when a rule is too vague to help the agent.

A useful test has:

1. a real brief
2. real sources
3. design constraints
4. a review step
5. a done condition

If you cannot say what "done" means, the agent cannot prove it.

## Save Failures Into The Harness

This is the ratchet.

When the agent fails, do not only fix the artifact. Fix the system.

Update:

- routing table
- `AGENTS.md`
- `DESIGN.md`
- skill instructions
- source map
- checklist
- failure log
- review process

The Harness gets better when mistakes stop being private lessons.

They become shared rules.

## Harness Artifact

### Agent Experience Audit

Use this as a first pass:

```text
Workflow:
Agent/task:
Human goal:
Sources used:
Context missing:
Design rules missed:
Assumptions made:
Validation run:
Human decision needed:
Failure class:
Harness update:
```

Start there.

One workflow. One failure. One permanent improvement.