Agent UI is a runtime-first draft standard for agent interaction surfaces. It defines how AI clients project structured agent events, tool calls, human-in-the-loop requests, artifacts, tasks, sessions, and evidence into user-visible UI without turning the UI into a second source of runtime truth.
Agent UI sits between runtime facts and product interaction. It does not own execution, context, storage, policy, or visual components; it defines how facts from those systems become visible, controllable, resumable, editable, and auditable.
| Adjacent system | It owns | Agent UI owns |
|---|---|---|
| Agent runtime | Runs, turns, tasks, events, snapshots, and authoritative status. | Projection into status, messages, tasks, and controls. |
| Tools and workflows | Executable actions, connectors, procedures, scripts, and tool results. | Tool progress, safe summaries, recovery states, and user control points. |
| Context and policy | Facts, sources, memory, permissions, risk, and trust boundaries. | Citations, missing-state UI, approval/input surfaces, and honest fallbacks. |
| Artifacts and evidence | Files, canvases, diffs, traces, reviews, replays, verification, and audit records. | Artifact cards, canvas entrypoints, timelines, evidence surfaces, and handoff states. |
| Design systems | Visual components, tokens, layout, and responsive behavior. | Surface semantics and behavior-level acceptance checks. |
- A runtime event projection contract for lifecycle, text, reasoning, tool, action, queue, artifact, evidence, and session events.
- Nine standard surfaces: Composer, Message Parts, Runtime Status, Tool UI, Human-in-the-loop, Task Capsule, Artifact Workspace, Timeline/Evidence, and Session/Tabs.
- A client implementation model for session hydration, progressive rendering, queue vs steer, durable snapshots, and controlled writes.
- Performance metrics for first status, first text, delta backlog, history hydration, timeline load, and artifact/evidence offload.
- Acceptance scenarios that prove behavior instead of checking whether a component or file exists.
agent runtime events + durable snapshots
-> projection reducer
-> UI projection store
-> Conversation / Process / Task / Artifact Workspace / Evidence surfaces
-> controlled runtime, artifact, and evidence actions
Compatible implementations should:
- Consume typed events and snapshots instead of parsing plain assistant text.
- Keep runtime, artifact, evidence, and UI projection state in separate owners.
- Show status before first text when the runtime has accepted work.
- Render text, reasoning, tool calls, action requests, artifacts, and evidence as different message parts or surfaces.
- Route approvals, interrupts, queue changes, steering, artifact edits, and evidence export through controlled APIs.
- Hydrate old sessions progressively: shell first, recent messages next, timeline/tool/artifact details on demand.
- Treat UI documentation as guidance only; no Markdown text is allowed to become policy, runtime state, or executable workflow.
Key pages:
- Specification
- Runtime standard
- Runtime event projection
- Session hydration
- Queue and steer
- Ecosystem boundaries
- 中文规范
- Agent Knowledge - source-grounded knowledge packs.
- Agent UI - interaction surfaces for agent products.
- Agent Runtime - execution facts, controls, tasks, tools, and recovery.
- Agent Evidence - evidence, provenance, verification, review, replay, and export.
- Agent Policy - policy decisions, approvals, permissions, risk, retention, waivers, and traces.
- Agent Artifact - durable deliverables, versions, parts, previews, exports, and handoff packages.
- Agent Tool - tool declarations, surfaces, invocations, progress, results, permissions, and audit refs.
- Agent Context - context surfaces, items, source refs, selection, budgets, assembly, injection, compaction, and missing-context facts.
See the Agent standards ecosystem page for the mutual-link map and future standard candidates.
npm install
npm run devnpm run buildThe static site is generated at docs/.vitepress/dist.