Activation Boundaries
Make the start, active scope, and end of ambient assistance explicit.
Reusable ways to make AI behavior understandable, useful, and controllable.
Make the start, active scope, and end of ambient assistance explicit.
Give each agent a stable identity, purpose, version, and visible permission boundary.
Help users evaluate agents by capability, permissions, provenance, and practical fit.
Coordinate agents through explicit ownership, handoffs, and shared-state boundaries.
Report useful agent outcomes with clear definitions and failure distributions.
Make agent updates traceable to a version and explain behavioral changes.
Expose context-relevant AI actions near a selection with a keyboard alternative.
Allow behavioral preferences while clearly separating tone from capability and permission.
Give contextual advice at the moment it improves an attempted task.
Use automated support for bounded tasks with a clear route to a person.
Provide quiet awareness of background work without implying constant recording or attention.
Enhance audio with a reversible original and an honest account of processing.
Summarize recordings with traceable timestamps and coverage limits.
Keep a legible event history for consequential AI actions with actors, times, and outcomes.
Show the chain of identities and permissions behind an agent action at a useful level of detail.
Suggest labels based on content while allowing correction and explaining scope.
Display active autonomous work, its current boundary, and how to intervene.
Bound unattended execution by understandable time, action, or spending limits.
Remove a background with visible transparency and edge inspection.
Process multiple inputs with per-item validation and visible partial outcomes.
Represent batch progress per item with cancellation, partial completion, and bounded retries.
Flag potentially unfair generalizations with explanation and a reviewable alternative.
Preview the people, records, systems, and costs an action may affect.
Explain when a result was reused and which inputs define its validity.
Offer a concise explanation of the approach and evidence rather than exposing private model reasoning or pretending it is an audit log.
Promote a useful deliverable into a focused workspace while preserving its conversation context.
Create a recovery point before a risky change and explain the scope of restoration.
Put verifiable references beside the claims they support, with a path from summary to source.
Provide a searchable action surface with meaningful names and keyboard access.
Organize component variations by meaningful properties and supported combinations.
Make conditional paths explicit with understandable criteria and default behavior.
Communicate uncertainty through evidence quality and meaningful categories rather than unsupported precision.
Show exactly which files or context items an AI request will use.
Let users explicitly reference an authorized entity in a request.
Move user-approved context between tools with clear contents and access boundaries.
Help people locate a past conversation using understandable query and filtering behavior.
Condense a long conversation while preserving decisions, unresolved questions, and provenance.
Organize conversations with editable labels whose meaning is understandable to the user.
Offer reusable conversation starting structures with editable assumptions and context.
Show operation cost in an understandable unit with estimate versus actual clearly separated.
Give people understandable access, export, and deletion controls for AI-related data.
Explain pricing changes in terms users can inspect and evaluate fairly.
Choose retry, correction, fallback, or stop based on the failure and side-effect risk.
Define when uncertainty, risk, or repeated failure should interrupt automation.
Export the intended conversation scope in a usable format with clear privacy expectations.
Explain a capability failure in plain language and offer an achievable next step.
Collect feedback tied to a particular result and explain what is shared.
Preview uploaded material and processing status before using it as AI context.
Offer relevant follow-up actions that reduce effort without constraining free input.
Warn about suspicious activity with a clear action and a path to correct false positives.
Generate charts from validated data with visible units and inspectable source values.
Generate interfaces from bounded components with predictable state and permission behavior.
Use gestures only when they are discoverable or accompanied by conventional controls.
Let people approve capabilities at the scope and moment where consequences are understandable.
Enforce a spending boundary across all relevant operations and expose the stop state.
Transfer a task to a human with preserved context and a clear status.
Put a person in control at a consequential decision with sufficient context to approve or reject.
Increase image resolution with clear source/output sizes and fidelity limits.
Edit a selected image region while preserving the unselected area and original asset.
Support spatial exploration with navigation, zoom, grouping, and an accessible alternative.
Attach a comment thread to a stable target with resolution and navigation.
Expose distinct input modes when the mode changes interpretation or required context.
Reduce purchase steps while retaining explicit price, recipient, payment, and final approval.
Translate a result while preserving access to the original and meaning-critical details.
Teach an interaction through a short achievable task with a clear exit.
Let people stop an incomplete response and continue from a clearly stated checkpoint.
Let users interrupt a speaking assistant through a simple accessible control.
Present inventory forecasts with uncertainty and the underlying planning window.
Expose useful relationships among retrieved entities without hiding the original source context.
Show collaboration presence only when it helps coordination and accurately reflects activity.
Display live transcription with recording state, speaker uncertainty, and correction.
Apply AI assistance to a selected region with a preview and preserved original.
Let people inspect, correct, and remove retained context with explicit scope.
Expose whether context is temporary, conversation-scoped, or retained beyond the session.
Let people keep a message visible as a reference without changing its underlying meaning.
Allow lightweight reactions associated with a specific message.
Help users choose a model by task-relevant trade-offs rather than unexplained internal names.
Use AI assistance to reduce form effort while keeping each step editable and validated.
Distinguish speakers and permission scopes on shared devices without assuming identity from a voice match alone.
Accept several input types with clear supported formats and privacy expectations.
Interpret a plain-language filter into visible editable constraints.
Remove selected visual objects while retaining an original and inspectable output.
Measure generated-output use with clear event meaning and privacy boundaries.
Let users choose an output structure before or after generation with clear conversion limits.
Keep a traceable history of generated outputs and intentional revisions.
Share only the intended result scope with clear access and revocation behavior.
Set permission at the level of an action rather than granting blanket autonomy.
Suggest performance changes from observed bottlenecks and explain their quality trade-offs.
Let users select a task-relevant perspective while labelling its limitations.
Adapt checkout to relevant known preferences while keeping all material choices visible.
Estimate likely operation cost with assumptions and a path to set a limit.
Suggest text ahead of the cursor while preserving authorship and edit control.
Let users create price alerts with scope, threshold, frequency, and a clear off switch.
Preview and control sensitive information before it leaves the user's boundary.
Set waiting expectations using measured ranges and visible uncertainty.
Show progress through real completed and current steps, with cancellation and a final outcome.
Keep the main result clear while letting people reveal detail when needed.
Reveal capabilities when context makes them useful without hiding essential controls.
Offer concrete starting requests that teach the input shape and product scope.
Translate a brief into an editable UI with transparent assumptions and real states.
Explain quota exhaustion with remaining allowance, reset timing, and viable alternatives.
Translate speech with visible language selection, delay, and access to original content.
Keep generated alternatives available for comparison with an explicit active version.
State what will be preserved, retried, corrected, and left unchanged during recovery.
Present resource use with understandable units, scope, and actionable thresholds.
Let people refine a specific result without losing the original or restarting the entire task.
Make it possible to distinguish human decisions from agent suggestions and tool execution.
Use return-risk estimates as decision support with fairness and review safeguards.
Label whether an action can be undone and what remains irreversible.
Summarize reviews with representative evidence and visible disagreement.
Show execution counters that map to real usage and remain secondary to the task.
Preview a proposed operation in an isolated environment before applying it.
State a recurring task's schedule, timezone, next run, and cancellation controls.
Help readers return to new output without forcing their scroll position.
Let an AI revise a suspected mistake using evidence and a visible revision boundary.
Retrieve by meaning while making query interpretation and source relevance inspectable.
Share a collaboration session with clear access scope and revocation.
Reserve plausible content structure while work loads without fabricating completion.
Offer structured commands inside an input without preventing ordinary language.
Offer completion suggestions without taking over typing or hiding the original input.
Recommend item combinations with explicit total cost and reasons for inclusion.
Present code with language, selection, copy, and execution boundaries clearly separated.
Compare options using consistent attributes and verified source dates.
Show proposed changes as reviewable differences before applying them.
Suggest form values from authorized context with provenance and review.
Offer recommendations based on relevant context with an understandable reason and correction path.
Provide an inspectable collection of retrieved documents linked to the relevant answer.
Reveal output progressively while keeping a clear stop action and final completion state.
Blend visual directions with interpretable parameters and comparable examples.
Apply a visual reference's broad properties while preserving subject and usage rights.
Distinguish proposing a plan, confirming an action, and executing it.
Expose image controls only when their effect and supported range are understandable.
Generate themes through semantic roles and preview real component states.
Let people explore an alternative from an earlier point without erasing the original path.
Allow a clear cancellation window before a deferred consequential action.
Offer a small set of understandable tone adjustments with an editable preview.
Let users change the tool or mode from the composer with visible scope and consequences.
Show which external capability is being used, what it returned, and whether it succeeded.
Publish understandable information about system behavior, limitations, and measured outcomes.
Combine items across sources while preserving merchant, delivery, price, and policy differences.
Guide people to a useful first task through a short set of relevant choices.
Present distinct generated alternatives with stable labels and an explicit selection.
Turn uncertainty into a concrete way to check the result before acting.
Search from an image with clear cropping, privacy, and similarity limitations.
Create synthetic voices only with appropriate rights, clear disclosure, and misuse controls.
Map speech to actions with visible interpretation and confirmation for consequences.
Confirm consequential spoken actions with an unambiguous summary and cancellation path.
Convert speech into an editable draft with clear recording start/stop and failure states.
Visualize real listening, thinking, and speaking states with non-motion labels.
Represent a sequence of steps and dependencies in a way people can inspect before running.
Connect claims to inspectable evidence without making the answer harder to read.
Keep people oriented while an agent searches, reads, or prepares an answer.