Agents you can see, steer and trust: the UX Room

UX work in a consultancy rarely fails loudly. It fails quietly: a brief nobody pinned down, research that sounds more certain than it is, a concept nobody can prove wrong. We built the UX Room to put AI exactly where those failures start, and to do it in the open.

The UX Room is a shared workspace with nine specialist agents: Orchestrator, Practice, Briefing, Research, Ideation panel, Concept design, UI & system, Critique and Prototype. Each has a narrow, written remit and a stated list of things it refuses to do. You choose who you talk to, you see how they hand work to each other, and every step is logged. It is deliberately not one assistant that answers everything.

For adesso, that distinction matters. Our clients do not buy output; they buy judgement they can rely on. This post explains why visible, steerable agents are the way to earn that trust, and what the approach can bring to our clients' projects.

The real risk with AI in design is not a wrong answer. It is a fluent, confident answer you cannot trace. In the UX Room, transparency is built into the behaviour of each agent rather than added as a disclaimer:

• Declared limits. Every agent states what it will not do. Research cannot browse or open files, so it says so and asks you to paste the source instead of inventing one. Critique never proposes alternatives. Prototype never edits the spec.

• Declared sources. Any document an agent writes opens with what it was built from and what was missing, so it cannot be read as more authoritative than it is.

• Honest about inputs. If you attach a screenshot the agents cannot see, they tell you plainly instead of pretending.

• A visible record. Every handoff and every decision you log lands in a timeline you can show a client.

A reviewer can ask of any statement: who said this, under which remit, and from what material? That is the standard we hold our own consultants to, and it is the standard we should expect from the tools they use.

Control starts with who you are talking to. You can simply describe the problem and the room routes you to one specialist (one, not a shortlist; if it lands wrong, that agent hands you on). Or you can pick someone yourself and open up to three agents side by side.

A few controls shape how they behave:

• Reference-only columns. Keep a second agent open to read alongside the one you are actually asking.

• Quick take or Probe first. Choose whether you want a best answer now with assumptions stated, or the one question that would change the answer.

• A shared case brief. Problem, who benefits, success criteria, constraints and scope are written once and read by all nine, so nobody has to re-explain the project.

• No forced consensus. When two agents disagree, the synthesis shows where and why. It does not average them, because averaging two specialists gives you a generalist.

The principle underneath: decisions that need a preference, such as priorities, trade-offs and whose interest wins, stay with the person. The agents set those decisions up; they do not take them.

Specialists are only useful if they respect each other's boundaries, so the room is built around handoffs rather than one agent doing everything. When a question is really someone else's, the agent gives its own partial angle, says whose it is and why, and offers a handoff.

The receiving agent does not start cold. It is told who sent the conversation, the reason, what you had asked, and what the sender had actually concluded. It opens by restating the question in its own words, so you can correct a handoff that lost something. The Orchestrator, who holds the process, also sees a digest of what the other agents have said, so you never re-brief the one agent whose job is to know the whole picture.

Two details make this safe for client work. Every handoff is offered to you and happens only when you take it, so a person stays in the loop at each transition. And every handoff is recorded in the Decisions and timeline view, with its reason, next to the decisions you chose to log. A Team view shows the same thing as a picture: which specialists are active, and which lines of work passed directly between them.

The gain is not speed for its own sake. It is catching expensive mistakes early, while they are still cheap to fix, and removing the re-explaining that eats a project team's week.

Take the moments of a typical project in turn. At the start of a case, the context is written once and read by all nine agents, so no one re-explains the project to each colleague or tool. During framing, Briefing asks who benefits and what each criterion rules out, which stops vague criteria that nobody owns from travelling downstream.

When choosing a direction, the Ideation panel spreads the options before anything narrows and never ranks them, so the choice stays with the team instead of settling on the first idea. In review, Critique names who fails, where, and what it costs them, instead of offering polite feedback that misses the blockers. And when work passes between roles, the handoff carries the reason and the conclusion and is logged, so the reasoning is not lost along the way.

AI will earn a lasting place in client work only if people can see what it is doing and stay in charge of what it decides. The UX Room is our working answer: specialists with clear boundaries, handoffs you can audit, and decisions that remain human. For adesso, that means offering clients pace and accountability together, which is where we can add the most worth.