Designing AI behavior, not AI screens

The screen is not the product

Open most AI product design files and you find screens. A prompt box, a results panel, a sparkle button, some empty states. All of it real work, none of it the actual product. With an AI agent the product is the behaviour: what it does with your input, when it acts on its own, when it stops and asks, what it says when it is not sure. You can lay out a beautiful screen on top of an agent that confidently lies, and you have shipped a beautiful problem.

So I treat behaviour as the design surface. The layout is downstream of it. Before I draw a single panel I want to know what the model does on its worst day, because that worst day is the one a user remembers. The work I lead on Articos, an AI user-research platform, pushed this from a slogan into a set of concrete moves.

Behaviour is a set of decisions you make for the model

Left alone, a language model has one default: sound confident. Designing its behaviour means deciding, in advance, where that default is wrong and what to do instead. A few of the calls I made on Articos:

• Cap confidence on unanimity. When every simulated voice agrees, a naive system reads that as certainty and turns the dial to the top. I treat agreement as a smell, not a signal, so the synthesis stage caps confidence at 75% when it detects unanimity. The behaviour is humility, and it is built in, not bolted on.
• Run a critique-repair loop. The model writes a section, then a separate stage critiques it against a cliché blocklist and a sycophancy check, then a repair stage rewrites the weak parts while keeping every source pin. The user reads the repaired version, not the first draft. The agent has already argued with itself before it speaks.
• Generate hypothesis-blind. The part that generates personas and scripts never sees the hypothesis under test. It gets role names and counts, nothing more. That single piece of withheld information blocks the most common path to a self-fulfilling answer.

None of those are screen choices. They are behaviour choices, and they decide whether the screen is worth building at all.

Guardrails are part of the behaviour, not a wrapper

A guardrail people picture as a fence around the model. On Articos the most useful guardrail is more like a vocabulary. Personas come from a curated library of lenses. The model picks a handful per study but cannot invent a lens outside the library. The library is the constraint, and because the lens names are stable, researchers build intuition across studies instead of relearning a fresh set of labels every time.

A good guardrail does not just stop the bad answer. It shapes the space the model is allowed to be creative inside.

That is the difference between a filter you apply after the fact and a behaviour you design into the generation step. The model still surprises you. It just surprises you inside a space you chose.

The agent proposes, the person decides

The behaviour I care most about is the handoff to the human. The model is allowed to be bold. It is not allowed to be the last word. On Articos the agent proposes a strong first draft and a person decides what to keep, at explicit checkpoints they own. I once shipped a one-click "generate full report" button because stakeholders wanted the demo. Every one of those reports got edited before anyone shared it. I killed the button and brought back the checkpoints, and in the next survey four of five researchers said the friction made them more confident in the output, not less.

So the rule is small and it does a lot of work: the AI proposes, the person decides. Most of the behaviour design is just making that handoff land at the right moment, with enough context for the person to actually judge it.

The harness is what makes the behaviour predictable

Behaviour you cannot reproduce is not design, it is luck. The thing that turns these moves into a product is the harness: the structure the model runs inside. On Articos the report pipeline breaks into stages, and each stage breaks further into more than a hundred substages, each one there to catch a specific failure the model produces when left alone. Theme drift, citation hallucination, overconfident synthesis, repair drift. Each has its own substage, its own guardrail, its own recovery path.

That harness is why the behaviour is predictable instead of vibes. The same study run twice goes through the same gates and gets caught by the same checks. A reviewer can point at exactly which substage capped a confidence score or rejected a theme. Designing AI behaviour, in the end, is designing that structure: the decisions, the guardrails, the handoff, and the harness that makes all of it repeat.

If you want the material underneath this, AI as a Design Material sets out the broader view: AI as a probabilistic raw material you design with, grain and all. And Building with LLMs goes inside the harness itself, the evals and routing and audit journals that hold the behaviour in place.