AI Agent UI Design: Dashboards, Approvals and Run History

AI agent UI design is the design of interfaces for agentic AI systems, products where the AI takes actions on behalf of users. This includes task queues, run history views, approval workflows, output review interfaces, and agent configuration screens. The design challenge is making autonomous actions feel controlled and auditable.

Transparency in agent UI means showing users what the agent is doing, has done, and is planning to do, in plain language. This requires clear activity logs, step-by-step action summaries, confidence indicators, and easy override or cancel controls. Trust is built through visibility, not just results.

Most agent products need: an agent configuration screen, a run view showing what the agent is doing right now, a history view of past runs and outcomes, an approval queue for actions requiring human sign-off, and an output review screen for reviewing and editing agent-produced content.

Agent errors are different from software errors; they're often ambiguous, partially completed, or the result of edge cases. We design error states that explain what happened in plain language, offer clear recovery paths, and don't require users to understand the underlying model to fix the problem.

Usually a combination, weighted by autonomy. Low-autonomy agents that answer questions suit chat. High-autonomy agents that act on their own suit an inbox-plus-dashboard model: a queue of completed and pending actions the user reviews, approves, or corrects, with chat available for instructions. The more the agent does alone, the more the UI should look like review tooling and the less it should look like a conversation.

The patterns we use most: a live run view showing each step as the agent works, an approval queue for actions above a risk threshold, activity logs written in plain language rather than tool-call names, confidence indicators on outputs, one-tap correction that feeds back into the next run, and a kill switch that is always visible. Together they make autonomy feel supervised instead of opaque.

A chatbot waits for a request and responds; the interface is the conversation. An agentic UI supervises work that happens without the user present, so it needs states a chat thread cannot express: queued, running, awaiting approval, partially failed, rolled back. Chat becomes one control surface among several rather than the whole interface.

Study products by the pattern they demonstrate, not the brand. For plan-first UIs, look at coding agents that propose an editable step list before running (Devin and GitHub Copilot's task workspaces popularized this). For live run views, Cursor's agent panel shows each action as it happens. For supervision at scale, agent platforms use an inbox of completed and pending runs. Our own shipped agent work, like the Echo AI dashboards in our portfolio, combines a run timeline with source-grounded outputs. We break these patterns down in our agent dashboard design guide.

Treat the plan as a contract. Before execution, show the proposed steps as an editable list so users can remove or reorder actions. During execution, mark each step as pending, running, done, or failed, with plain-language labels instead of tool names. After execution, show what changed with a diff or summary per step. Plans above a risk threshold should require explicit approval before the agent starts.

The decisions that happen before any screen: which actions the agent may take alone and which need approval, how risk thresholds are set, what the user is told when a run fails halfway, how they are notified without being flooded, and how work hands back from agent to human. Get those rules right and the UI mostly follows. Get them wrong and no amount of visual polish makes the agent feel safe to use.

Our subscription starts at $4,917/mo. AI agent design projects start from $3,000 for one-off engagements.