The sustained mediation between actor intent and system capability, transforming fuzzy goals into concrete outcomes through an adaptable intermediary that evolves with use. This umbrella term encompasses all user-facing AI interactions — from the strict sense of a model running tools in a loop to achieve a goal, to chat interfaces where tool use is incidental or absent.
Agency
Agent patterns embody different levels of system agency from the agency spectrum:
- Chatbot: User-driven interaction requiring explicit prompts
- Inline: Assistance mode providing contextual help and smart suggestions
- Workflow automation: autonomous action that follows predefined sequence
- Full autonomy: proactive monitoring and acting without direct user input
Higher agency modes require more careful preservation of user agency to maintain control and transparency.
Modes of interaction
- Focused: Agent as a primary interface, typically via messaging.
- Inline: Contextual interactions embedded directly in the UI, such as AI completion, rewrite suggestions, or smart defaults.
- Ambient: Behind-the-scenes monitoring or support, offering proactive help or taking actions without direct prompting. Next-best action is the primary pattern here.
Variants
Single-turn interactions
Simple query-response patterns without state persistence or tool use.
Multi-turn conversations
Maintaining context across exchanges, allowing for clarification, follow-ups, and evolving understanding.
Tool augmentation
Agents that can invoke external capabilities—from simple calculations to complex API calls. This includes the bounded loop execution pattern where agents run tools iteratively until reaching objectives.
Memory
Short-term
Conversation context provides implicit short-term memory, allowing agents to maintain state across interactions within a session.
Long-term
Persistent memory requires explicit implementation through:
- Storage tools (databases, files, vector stores)
- Retrieval mechanisms with semantic search
- Memory management strategies to prevent unbounded growth
Anti-patterns
Algorithmic loafing
When agents take creative or cognitive lead, humans reduce effort and disengage—accepting agent output with minimal reflection. This manifests as:
- Reduced cognitive engagement despite task importance
- Loss of skill development and expertise over time
- Decreased perceived ownership and commitment to outcomes
- Homogenisation of outputs across users
The effect parallels social loafing, where individuals reduce contribution when responsibility diffuses.
Mitigation: Design interactions where humans remain in creative or cognitive lead whilst agent provides scaffolding through questions, analogies, or contextual knowledge rather than complete solutions. See Conversation dialogue forms for question-driven approaches that maintain engagement.
Mismatched scope
Applying agents to tasks requiring genuine accountability, sustained comprehension, or human empathy. Agents cannot take responsibility for actions or form independent intentions, and their understanding operates within learned patterns without true comprehension.
Uncalibrated trust
Treating agent outputs as authoritative without verification mechanisms. Agents can generate plausible but incorrect information, and performance degrades with extended interactions or complex state.
Undefined boundaries
Failing to communicate agent capabilities and limitations, leading to misuse and frustration. Users need clear understanding of what agents can and cannot do to set appropriate expectations.
Related components
Resources & references
- Maier, Schneider, Feuerriegel (2025) Partnering with Generative AI: Experimental evaluation of human-led and model-led interaction in human-AI co-creation
- Nardi, B., Miller, J. R., & Wright, D. J. (1998) Collaborative, programmable intelligent agents. Communications of the ACM, 41(3), 96–104.
- Thu, Kocaballi (2025) Personalization features for Human-LLM Interactions
- Simon Willison (2025) Agents
- Amelia Wattenberger (2024) Why chatbots are not the future
- Luke Wroblewski (2023) Expanding conversational user interfaces
- Andy Matuschak (2025) How might we learn? – guidance in action and the critique of transactional chatbot tutoring
Consequences
- there is a standing intermediary between intent and capability: the actor says what they want and something else decides how, which is both the whole benefit and the whole exposure
- how much freedom the agent has to act unprompted is a setting rather than a fixed property, so the actor's trust has to be calibrated to whatever it currently is — and they have to be able to find that out
- the work happened somewhere the actor wasn't watching, so it owes an account of itself afterwards to be reviewable at all
Related patterns
Precedes
- Activity log — actions and reasoning are tracked and displayed in the activity log.
- AI tuning — tuning keeps the actor in control while the agent does its thing
- Generated content — presentation and manipulation of AI's output.
- Transparent reasoning — generates the reasoning that transparent reasoning makes visible and inspectable.
- Live presentation — generates the stream that this surface presents in flight
Instantiates
- Assistance — the chatbot and inline agent modes apply the assistance configuration
- Delegation — the ambient and workflow automation agent modes apply the delegation configuration
Enacts
- Agency — each mode hands the system a different amount of freedom to act unprompted
- Formality — agents accept unstructured input by design; structure can be inferred or gradually formalised from interaction.
- Learnability — the agent lets the actor ask the system rather than decipher it; learning happens through dialogue at the actor's pace
- Adaptability — the agent modulates its behaviour and output to context and inferred intent rather than following fixed rules
Complements
- Collaboration — Agents participate as collaborators with adaptable mediation capabilities
- AI completion — powers AI completion via the inline mode of interaction
- Next-best action — powers next-best action via the ambient mode of interaction
- Workflow — gives the workflow-automation agent mode its procedure — the agent runs it, this artifact shows and edits it
Tangentially related
- Localization — multilingual interaction
Alternatives
- Wizard — can guide through complex processes with conversational flexibility wizards lack
Related
- Assisted task completion — powers AI completion and next-best action; the inline and ambient modes map directly to this group
- Capability & scope — the conversational partner whose capabilities are being stated
- Data entry — the agent often responsible for interpreting forgiving input.
- Help
Preceded by
- Onboarding — the first steps that help people get to know the agent — what it can and can't do
- Prompt — the prompt articulates the request the agent acts on
Enabled by
- Conversation — messaging is the primary interface for conducting agent conversations