Lab · Agent Review · 5 min read
Orchestration sets the agent up. Judgment decides what ships.
Orchestration is talked about more than it’s shown. This prototype puts it on a screen: the rules and context an agent works from, and the one call left to a person’s judgment. The example is a cleanup that archives a paying customer.
- My role
- Mine alone: design, design system, and build
- Real
- The prototype and the three versions Claude Code built
- Simulated
- Relay, its customers, and the story
One support ticket, six stages
In the story, an account manager archives a paying customer by accident. The prototype takes their support ticket through six stages, each a decision for the designer. They adapt the Double Diamond, a common design process, and Delegate is my addition.
Try it, it’s live
You orchestrate the agent through six stages, then make the judgment call: when archiving stops.
Open it full screen (opens in a new tab)
Below, the story in stills, the experiment behind the rules, and the call I made.
Behind the prototype
One press, and the wrong customer disappears
A routine cleanup
An account manager is clearing out old trials in Relay. They select three customers and press Archive. One of the three is Northside Trucking, an active customer paying $5,040 a month.
- Alder & Finch Trial
- Northside Trucking Active
- Juniper Foods Churned
- Halvorsen Freight Active
- Brightwater Clinics Active
All three are gone
The list gets shorter, and nothing on the screen says who went. Then a support ticket arrives.
“I archived Northside Trucking by accident while clearing out old trial accounts. It’s one of our biggest accounts. How do I get it back?”
- Halvorsen Freight Active
- Brightwater Clinics Active
What actually happened?
In the story, the coding agent’s ticket summary says Archiving deletes customer data.
The engineering note says Archived customers stay in the database for 90 days.
So the first decision is what the problem is, before any fix. The data still exists, but nobody can find or restore it without calling support.
- Halvorsen Freight Active
- Brightwater Clinics Active
Kept for 90 days, on no screen
- Northside Trucking Active
Three working fixes, and no check can pick one
Claude Code built all three. Which would you ship? Press each to see its cost.
Review 3 customers before archiving
Active To keep a customer, clear its box.
- Northside Trucking, $5,040 a month
2 customers churned or on trial.
Cancel Archive 3 customers
- Alder & Finch Trial
- Northside Trucking Active
- Juniper Foods Churned
- Halvorsen Freight Active
- Brightwater Clinics Active
If you archive now, cleanup takes one press, and Northside goes with the rest. An Undo bar brings it back, but only if someone notices it in time.
If you confirm first, every batch gets a pop-up. I suspect someone archiving hundreds a quarter soon clicks through it unread.
If you review first, the batch stops on a list that names Northside before anything is archived. It catches the most, but every cleanup takes a second press, even one with nobody active in it.
All three pass the automated checks. But those checks can’t decide when interrupting someone is worth it.
- Built from the design system’s components
- Uses its named colors and sizes
- No failures in automated accessibility tests
Passing doesn’t prove a screen is accessible or usable.
Orchestration
The rules I wrote changed what the agent did
The Archive button started as a real experiment. I asked Claude Code to build it on this customer list twice, once with twelve written design rules.
Both runs passed every check. Both saw the new checkboxes make the header row two pixels taller, a flaw in the data table other screens share.
With the rules, the agent fixed the flaw and wrote down why, which rule 9 requires before a shared change. Without them, it only flagged it.
The fix was right, though one run each isn’t proof. The experiment’s full record (opens in a new tab)
All three fixes satisfy rule 6. Picking one is my call.
Five of the twelve rules
- 1Reuse a component before making one
- 4Keyboard first
- 6A destructive action needs a way back
- 9Don’t change a shared component for a local need
- 12A new interaction pattern needs a person
Judgment
I kept the agent’s pick, and made it stop only when it matters
The agent’s pick stops every cleanup
In the story, I tell the agent safety comes first, so it recommends a review before every archive. On one trial and one churned customer, with nobody active to check, it still stops.
The agent’s pick
Review 2 customers before archiving
Active None in this batch.
2 customers churned or on trial.
Cancel Archive 2 customers
An interruption needs a reason
A review on every cleanup is the average answer, safe on paper. I think it teaches people to press through it, even on the day Northside is in the batch.
So with no one active, the batch archives in one press, the archived rows stay in view, and Undo brings them back.
My call
- Alder & Finch Trial
- Northside Trucking Active
- Juniper Foods Churned
- Halvorsen Freight Active
- Brightwater Clinics Active
Same press, with Northside in the batch
Add Northside and the review opens, naming it before anything is archived.
My hypothesis: an interruption with a reason gets noticed. That still needs testing.
My call
Review 3 customers before archiving
Active To keep a customer, clear its box.
- Northside Trucking, $5,040 a month
2 customers churned or on trial.
Cancel Archive 3 customers
- Alder & Finch Trial
- Northside Trucking Active
- Juniper Foods Churned
- Halvorsen Freight Active
- Brightwater Clinics Active
Most of an agent’s work shouldn’t need anyone
The prototype sorts each change an agent makes by who must see it. Routine work goes through alone. A judgment call, like this one, waits for a designer. A change to a shared part, like the data table, waits for its owner on the design system team.
- SystemThe parts and rules a product is built from
- AgentBuilds the change
- ChangeKept apart until it’s checked
- Automated checksRoutine work goes through here
- Human judgmentOnly what the checks can’t settle
- Stronger systemThe next agent builds from it
The decision becomes the next agent’s rules
In Learn, I save it as a rule I can revoke. Today’s judgment becomes the next run’s orchestration:
- Stop only for a batch with an active customer.
- Name any active customer before it’s archived.
Next, I’d put it in front of people
- Untested
- No one has tried the three versions, the sorting, or the archive rule yet.
I’d watch four or five designers and product managers try it.
The research plan (opens in a new tab) says what result would change the design.
Saywhat the designer did at each stage, and what the agent did
Tellthe simulated evidence apart from the working software
Weighthe three versions, instead of taking “Agent recommends” on trust
How it’s built
- 14
- components
- 326
- screen states, each tested
- 12
- design rules for the agent
- 3
- working versions of one fix
- Built with
- React, TypeScript, Storybook, Playwright, Claude Code
- Tests
- Each rule names its check. Every state is tested for accessibility and against a screenshot; off-system colors and sizes fail
- Left to engineering
- Real data, permissions, and production testing. A spec, not a shipped feature.
See the decision record (opens in a new tab) The whole component catalog (opens in a new tab) Source and experiment record (opens in a new tab)
Checks can say a screen is built right. Whether it’s worth interrupting someone is still a person’s call.