1Overview

Agents scale.
A person’s attention doesn’t.

Problem

Engineers hand tasks to coding agents, step away, and come back to code they didn’t watch being written. The engineers I talked to were already running several at once across branches, so the work had become managing the agents.

Design question

How might one engineer stay in charge when agents write more than anyone can read?

Key insight

Most agent changes don’t need approval. Another agent can verify them, or someone just needs to know they happened.

What I designed

The concept and the prototype: the track that shows every agent at once, the calls a person answers, and a Lock Screen version for when they step away.

2Process

It started as an IDE extension for one agent.

First interviews

I interviewed four engineers. They didn’t trust an agent’s work until they had checked it themselves, by rerunning it, restating what they asked for and getting a teammate to look. So I started with something closer to an IDE extension than a standalone product.

Version one

5 screens

A review panel that checks an agent’s work against the project spec. A staff engineer told me most work starts from a casual request, and a formal spec is rare.

Version two

6 screens

A guided view of what the agent did while you were away. An industry research engineer told me they actually run several agents at once, and the staff engineer had said the same. That was the turning point.

With Claude Code

2 screens

I gave it all my research and asked about running many agents. It brought a lot of new features, but the design wasn’t consistent, and the solution looked like AI slop you could see anywhere.

Paper sketches

I drew how a race would work, listed what the design still had to answer, and wrote rules for prompting.

Version three

Play

That became the track: every agent is a lane on one screen, and a baton goes up when one needs you. A process validation engineer asked for two more things: every conclusion traces back to its source, and every decision shows what it will affect.

3Orchestration

Agents pass the work along like a relay.

4Decisions

The hard part is how often agents should ask.

5Management

Once agents are running, the job looks like a manager’s.

6Compared

Today’s tools run agents in parallel, then hand the work back to you.

Triage

Claude Code’s agent view is the closest to what I wanted: sessions sorted by what needs input, with the question on the row. It can’t say why this one needs you, or what’s waiting on it. In Relay a call comes with its reason, and the one that frees the most lanes comes first.

Output review

Cursor’s Agents Window gives you the agent’s summary next to the code. That’s the agent describing its own work, and you still have to picture what changed for the customer. Relay shows each fix the way it would ship. Try it

Mobile

GitHub Mobile’s Live Activity tells you how a session is going, and Claude Code’s Remote Control lets you answer from the Claude app. I wanted the question already in the notification, with its answers one tap away.

7Results

Relay goes on show at the Fall Symposium.

Feedback

A staff engineer gave the first version high praise compared with their company’s own internal tool.

What I learned

Both premises that fell over were mine, and each fell the first time an engineer saw the design. Next time I’d ask for a real example of their day before showing a screen.

With Claude Code, the features came fast and none of the screens were ready to use. It got useful once the research, the sketches and the rules came first.

What I’d watch first

How often people open the evidence before taking a recommended answer. When that drops, they’ve started approving without reading.