Framework 05 of 12 · Active operation Phase · Oversight triggers
Interrupt Pattern Library
The art of knowing when to speak and when to stay silent
A catalogue of interrupt patterns for agentic systems - when to surface decisions to humans, how to present them, and how to calibrate frequency against consequence. The art of knowing when to speak and when to stay silent.
Interrupt Pattern Library: Core Principles
Every Interrupt Has a Cost
Each time the agent interrupts the human, it imposes a cost: context switching, decision fatigue, and erosion of the delegation value proposition. The Interrupt Pattern Library treats interrupts as a scarce resource to be spent wisely, not a default behaviour to be applied liberally.
Consequence Determines Interrupt Priority
The decision to interrupt should be proportional to the consequence of not interrupting. A price change within normal range does not warrant interruption. A vendor going out of business mid-order does. The library provides a consequence taxonomy that maps situations to interrupt urgency levels.
Context Must Be Preserved Across Interrupts
When the agent interrupts the human, the human needs context to make a decision. The interrupt must carry enough information for the human to understand the situation, evaluate options, and respond - without requiring them to reconstruct the entire history of the agent's operation.
Batch When Possible, Interrupt When Necessary
Multiple low-consequence decisions should be batched into periodic summaries rather than delivered as individual interrupts. Only time-sensitive or high-consequence situations warrant real-time interruption. The library defines batching strategies that reduce interrupt frequency without sacrificing decision quality.
The Human's State Matters
The same situation might warrant different interrupt approaches depending on the human's current state. During work hours, a brief notification might suffice. During sleep, only critical interrupts should break through. The library includes patterns for state-aware interrupt delivery.
The value of an agentic system is measured not by how much it does, but by how little it needs to ask. Every unnecessary interrupt is a failure of design - evidence that the system could not handle what it was delegated to handle.
Interrupt Pattern Library: Implementation Patterns
Interrupt Taxonomy by Consequence
A classification system that maps situations to interrupt levels: informational (can wait), advisory (should see soon), urgent (needs attention now), and critical (must act immediately). Each level has defined delivery channels, persistence rules, and escalation timelines.
When to use: When designing the interrupt strategy for any agentic system.
Presentation Modalities
Interface patterns for different interrupt types: ambient indicators for informational updates, notification cards for advisory items, modal dialogs for urgent decisions, and multi-channel alerts for critical situations. Each modality is designed to match the urgency of the situation without creating alarm fatigue.
When to use: When implementing the delivery mechanism for each interrupt level.
Frequency Calibration
Algorithms and heuristics for managing interrupt frequency over time. Includes patterns for reducing interrupts as the agent learns user preferences, increasing interrupts when entering unfamiliar territory, and adapting to user feedback about interrupt volume.
When to use: As a continuous optimisation process running alongside the interrupt system.
Context-Preserving Interrupts
Design patterns that package sufficient context with each interrupt for the human to make an informed decision without returning to the full application. Includes situation summaries, decision options with trade-offs, recommended actions with rationale, and one-tap response patterns.
When to use: For every interrupt that requires a human decision or acknowledgment.
Batch vs. Real-Time Patterns
Decision frameworks for determining whether a situation warrants real-time interruption or can be included in a periodic batch summary. Includes time-sensitivity assessment, consequence decay analysis, and user preference integration.
When to use: When the agent encounters a situation that might warrant human attention but is not clearly urgent.
Interrupt Pattern Library: Commerce Applications
Exception Escalation in Agentic Shopping
When the machine customer encounters an exception during agentic shopping - an item out of stock, a price above the budget ceiling, a delivery delay beyond the specified window - the Interrupt Pattern Library determines whether to escalate to the human, substitute autonomously, or pause the transaction. Each exception type maps to an interrupt level based on consequence and reversibility.
Price Alert Calibration
In dynamic pricing environments, the agent must decide which price changes warrant human notification. A 2% fluctuation on a routine purchase is noise. A 40% spike on a time-sensitive purchase is critical. The library provides price-sensitivity models that calibrate alerts to the specific transaction context and user preferences.
Vendor Communication Relay
When vendors communicate during a transaction - shipping updates, substitution offers, delay notifications - the agent must decide what to relay to the human and what to handle autonomously. The library provides filtering patterns that surface meaningful updates while absorbing routine operational noise.
Multi-Order Coordination
When the agent manages multiple concurrent purchases, interrupt batching becomes essential. Rather than notifying the human about each order individually, the library provides patterns for consolidating updates into coherent summaries that give the human a complete picture without overwhelming them.
The worst interrupt is the one that should have happened but did not. The second worst is the one that happened but should not have. The Interrupt Pattern Library navigates between these failures with consequence-calibrated precision.
Interrupt Pattern Library: Guidance for Teams
Start With
- -Catalogue every situation where your agent currently interrupts the human
- -Classify each interrupt by consequence level (informational/advisory/urgent/critical)
- -Identify interrupts that could be batched into periodic summaries
- -Build context-preserving templates for your top 3 interrupt types
Build Toward
- -Machine learning models that predict optimal interrupt timing from user behaviour
- -Cross-channel interrupt coordination for users across multiple devices
- -Organisational interrupt policies that set maximum frequency thresholds
- -A/B testing frameworks for interrupt presentation and timing
Measure By
- -Interrupt-to-action ratio - how often do interrupts lead to meaningful human decisions?
- -False alarm rate - how often do interrupts prove unnecessary after the fact?
- -Response latency - how quickly do users respond to different interrupt levels?
- -User satisfaction with interrupt frequency and relevance
Interrupt Pattern Library: Lifecycle Connections
Framework 03
Autonomy Gradient Design System
Lower autonomy levels require more interrupts. The Autonomy Gradient determines the baseline interrupt frequency.
Explore frameworkFramework 06
Multi-Agent Orchestration Visibility
In multi-agent systems, interrupts must be coordinated across agents to prevent overwhelming the user.
Explore frameworkFramework 09
Explainability & Observability
Every interrupt is an opportunity for explanation. Explainability patterns determine how much reasoning context accompanies each interrupt.
Explore frameworkInterrupt Pattern Library: What Comes Next
Interrupt patterns govern single-agent communication. The next framework - Orchestration Visibility - addresses how humans maintain oversight when multiple agents collaborate.
Interrupt Pattern Library: The Framework Ecosystem
Navigate the complete lifecycle of Agentic Experience Design. Each framework addresses a distinct phase of the human-agent relationship.