← Selected work

Case Study — 02 / AI Agents

Autopilots

A framework and prototype set for turning single-purpose agents into teammates — with identity, initiative, shared context, and a trail people can actually inspect.

Role

AI Product Design

Year

2026

Tools

Figma, GitHub Copilot CLI, Claude Code

Prototype

Try it yourself ↗

01 — Context

The gap

Most AI agents can answer a question. Almost none of them can hold a job.

Ask, answer, done. That loop is useful, but it can't carry work that runs across days, people, and systems. So I started with a simple question: what would it take for an agent to earn a seat on the team?

 
Agents today
A teammate
Identity
Forgets you after the last message
Knows who you are and what you are working on
Initiative
Waits to be asked
Notices what changed and acts first
Teamwork
Works alone
Hands off to other agents like a team would
Paper trail
Asks you to trust it
Shows what it did, and why

So what

The gap isn't intelligence. It's everything around the intelligence — memory, initiative, handoffs, and proof.

02 — Framework

Four qualities

Four qualities separate a tool you use from a teammate you trust.

It knows who it is. It speaks up first. It works with other agents. It shows its work. Each one costs something real to build, and none of them work on their own.

01
Identity
It knows who you are and what you are working on — past the last message.
The most expensive to build. It needs a real profile, not a chat history.
02
Initiative
It notices what changed and starts the work before you ask.
Useless without identity. Acting on the wrong context is worse than waiting.
03
Teamwork
It hands work to other agents the way a team hands off between people.
Needs shared context. Without it you just get more silos, running faster.
04
Paper trail
It shows what it did and why, in a form you can actually check.
The cheapest to build. Buys the most trust.

So what

Showing its work is the cheapest to build and earns the most trust. Identity is the most expensive — and everything else leans on it.

03 — Proof

Scenario 1 of 3

When something breaks, the real work is coordination — not answers.

When a store is suddenly short-staffed, the autopilot recognizes the staffing problem, pulls the right context, recommends coverage, and keeps the manager in control instead of making them chase status.

Open this scenario full screen ↗

So what

One agent held the thread across six tools. The people made decisions instead of chasing status.

03 — Proof

Scenario 2 of 3

A new hire should arrive to a desk that is already set up.

The moment someone is hired, the same agents coordinate setup, a tailored first week, and a briefing for their manager — so everything is ready before the new person walks in.

Open this scenario full screen ↗

So what

None of this needed new technology. It's information the company already had, gathered by something that understands the job.

03 — Proof

Scenario 3 of 3

When the same question keeps coming back, the pattern matters more than the replies.

When the same kind of question keeps spiking, the system treats it as a signal rather than more work. It adapts its answers, and it tells the people who should know.

Open this scenario full screen ↗
What the three scenarios covered
 
Identity
Initiative
Teamwork
Paper trail
Crisis response
A new hire
A spike in questions
Fully exercised Partly exercised

So what

Noticing the pattern is the work. Answering the questions one at a time is just triage.

04 — Risk

Sequencing

Build these in the wrong order and you ship something worse than nothing.

The four qualities only work together. On its own, each one fails in a way people spot straight away.

Initiative without identity
Surveillance
Teamwork without a paper trail
A black box
Identity without initiative
Just a login
Any of it without teamwork
One more silo

So what

Identity first, then the paper trail. Initiative before either of those is just software watching you work.

05 — Craft

How it was made

I built the working prototype myself, in about a week.

I wrote it first, then designed it, then built it — handing each step to whichever tool was best at that part.

STEP 01
A written brief
The argument first, in plain prose. If it didn't hold up as writing, it wasn't going to hold up as a screen.
STEP 02
Figma
Shapes, flows and states — fast and rough, just enough to see the shape of it.
STEP 03
An agent wired to the design file
A coding agent reading the Figma file directly, turning frames into real components.
STEP 04
A working prototype
Clickable, end to end — the same one embedded at the top of this page.

So what

Being able to build the thing changed the argument. People stopped debating the idea and started using it.

06 — Reflection

What I learned

The AI was not the hard part; trust was.

It was deciding what a teammate should actually feel like. The four qualities became my test for every decision: if a feature didn't add identity, initiative, teamwork, or a trail, it didn't belong. That one constraint is why the prototypes took weeks instead of months.