Design Outcomes Weekly reflections on the real-time work of supporting a design org.
Johannes Vermeer’s Woman Holding a Balance: a woman in a white head scarf and a blue fur-trimmed jacket stands at a table by a window, holding up a small balance with empty pans, pearls and gold spilling from a box beside her and a painting of the Last Judgment on the wall behind her.

She checks the balance before anything goes on the scale, the way we want our agents to.

Why this matters

An agent earns trust the same way a good teammate does: it knows what it can take, it checks before touching other people’s work, and it reports back on what it couldn’t do. Here’s where that showed up 3 times in one week.

Johannes Vermeer, Woman Holding a Balance, c. 1664. National Gallery of Art, Washington

An Agent Should Know What It Can Take Two mindsets for agents in a product, the triage pattern that decides what an agent takes on by itself, and the report-back moment that makes automation trustworthy. An agent earns trust the same way a good teammate does: it knows what it can take, it checks before touching other people’s work, and it reports back on what it couldn’t do. Here’s where that showed up 3 times in one week. Give every agent 3 moves: take the work it’s confident in, check with the people its work touches, and hand off what it shouldn’t do with full context. Then design the report-back screen first, because the handoff is where automation earns the right to act.

An Agent Should Know What It Can Take

Two mindsets for agents in a product, the triage pattern that decides what an agent takes on by itself, and the report-back moment that makes automation trustworthy.

How much should an agent do on its own? The question came up in 3 different venues this week, and the same answer kept coming up: the design work lives at the boundary. An agent needs to know what it can take, what it should check before touching, and how to report back on what it couldn’t do.

The first room was a 1:1 about agents on our billing surfaces, where one of our design managers and I landed on 2 mindsets an agent can serve. There’s a collection mindset: right after a session, something timely starts the follow-up on payment, and it doesn’t even need to live inside the product to do it. And there’s a reconciliation mindset: at the end of the day, the agent gives you an update on what it did for you, highlights what you need to go do because it can’t, and answers questions about any of it along the way. The manager’s version of the second one is an end-of-day assistant that closes out your day and shuts your books down for you. What I like most is where the idea came from. The team had spent cycles polishing a predictive feature that can never be fully accurate, hoping it would change behavior, and the manager’s reframe was to stop predicting: find out what the behaviors are, and build product around the behaviors everybody is already doing. The behavior, it turns out, is settling your work right after the session or at the close of the day. So that’s where the agents go.

The second room was about our own internal workflow. One of our engineering leaders is building a system where every piece of customer feedback lands in one place, with agents on top of it, and the operating rule is the cleanest statement of agent boundaries I heard all week: if there’s anything high confidence that can get fixed because it was flagged somewhere, it’ll just get fixed. It’ll get done. Everything that needs more human judgment gets pushed to triage instead. A feedback repository stops being an archive you visit and becomes a work queue that drains itself.

The triage layer is where the manners live.

The agent will look at it and say: yep, this is something I can take, let me run it all the way to review. Or: here are the people this touches, let me message them first. And if it’s something the agent shouldn’t do, it hands the work off with all the right context.

Those are the same manners we’d ask of a new teammate: take what you’re sure of, check with the people your work touches, and pass along what isn’t yours with enough context that the next person doesn’t start from zero.

The third room is the reason this matters to the people we design for. Clinicians have been wary of automation since long before AI: anything that acts on their behalf has to leave them in control. That’s why the reconciliation report is the screen I’d design first on any agent, before the clever parts. The moment where the agent says what it did, what it couldn’t do and what it needs from you is the moment it earns the right to act at all. If your team is putting an agent into anything, I’d start the design there, at the handoff, and let the autonomy grow only as fast as that report stays trustworthy.

Filed under TR The Reframe — Assumptions, examined.