Agents: Make tools, approval, budgets, and rollback visible before agents act.
Essential
More in Agents
- Per-Action AutonomyAutonomy scoped per capability, not per app
- Checkpoints and RestoreNamed snapshots with one-click restore
- Workflow BuilderVisual drag-and-drop workflows
- Memory ManagementSee, edit, and delete what the assistant remembers
- Human HandoffEscalate from the agent to a human with context
- Escalation ThresholdsAuto-demote autonomy when risk crosses a line
- Error Recovery StrategiesConfigurable retry/fallback patterns
- Self-CorrectionAgent detects errors and retries with a visible fix
- Conditional Logic & BranchingBranch workflows by conditions
- Scheduled Tasks & Recurring ActionsTime-based automation triggers
- Multi-step Forms with AIAdaptive progressive forms
- Agent OrchestrationVisual flow for multiple agents
- Agent Performance MetricsDashboard showing agent success rates
- Agent VersioningA/B test different agent configurations
- Agent MarketplaceBrowse and install pre-built agents
- Blast Radius VisualizationPreview scope before agent executes
- Reversibility MarkingLabel actions by how easily undone
- Time-Delayed ExecutionCancelable countdown before high-impact actions
- Pre-Task Cost EstimateForecast tokens, dollars, and time before run
- Autonomous Mode DisplayPersistent signal when agent runs unattended
- Context PortabilityLegible payloads across app boundaries
- Ambient Presence DisplaysPeripheral signals for agent listening or idle
- Sandbox PreviewDry-run plan with explicit side effects before execute
Frequently asked questions
How are agent patterns different from chatbot patterns?
Chat patterns focus on conversation contract and turns. Agent patterns focus on side effects, tool calls, permissions, checkpoints, and durable control across sessions.
What should users always see before an agent acts?
Intent, scope, and reversibility: which tools will run, what data they touch, and how to stop or undo. Approval workflows and autonomy budgets make that contract explicit.
When do I need a task queue versus suggest / confirm / execute?
Use suggest / confirm / execute for the session-level autonomy contract. Expose a task queue only when several agent jobs run in parallel over time and users need status per job—not as a separate catalog pattern from tool use.
How do I design rollback and checkpoints?
Expose checkpoints before destructive steps, label what state will be restored, and keep rollback one gesture away. Users forgive agent mistakes when recovery is obvious.
Which products exemplify agent UX in this catalog?
Pattern examples reference Cursor, Devin-style agents, workflow builders, and assistants with tool pickers. Open any card’s examples list for specific references.
Where is the Agentic UX framework?
Use the linked framework for territories (permission, rollback, autonomy) and anti-patterns. Category patterns here are the implementation catalog mapped to those territories.