Pattern library
Agent design patterns
How a person hands work to an AI agent, stays in the loop while it runs, checks what was done, and supervises several agents at once. Seven families of interaction pattern, drawn from shipped products and published research.
What changes when the loop opens
An AI agent changes what an interface has to do. A conversational assistant answers inside one sitting, with the person present and judging. An agent works after the person has closed the window, takes actions in systems the person cannot see, and sometimes splits the job across several agents at once.
The patterns below describe how teams handle that. They are drawn from products shipped by Google, Anthropic, Microsoft, LangChain, Asana and Notion, and from human-computer interaction research published in 2025 and 2026. Each one names the product or the paper it came from.
Emily Campbell’s Shape of AI is the reference pattern library for AI interfaces and covers the prompt-and-response loop in depth. This resource covers the separate problem that appears once that loop opens.
- Time: the work outlives the session. A research run, a migration, an overnight job. The person who briefed it is not there when it finishes.
- Plurality: one request becomes several agents, and one agent serves several people. Neither fits a single conversation thread.
- Initiative: the system starts work in response to an event or a schedule rather than in response to a prompt.
- Distance: the work happens in a sandbox, a browser, or another company’s API, where the person cannot watch it.
The handover cycle
An agentic task moves through six stages, and control sits in a different place at each one. The person sets the brief. The agent drafts a plan and the person edits it. The agent runs. Both parties meet at checkpoints. The person verifies the result. What the system learns becomes the next brief.
Products built in 2025 and 2026 are consistently strong at the run stage and thin at brief, verify and learn. That imbalance is where delegation stops, because a person who cannot state the boundaries up front and cannot check the work afterwards will not hand over a second task.
The seven families
| Family | The question it answers | Patterns |
|---|---|---|
| Handover | What exactly was asked for, and what is the agent allowed to do? | Intent preview, editable plan, clarifying interview, goal and task separation, autonomy dial, budget envelope, durable briefs, plan learning |
| Oversight in motion | How does a person stay in the loop without living in it? | Scheduled confirmation, action guards, take over and hand back, mid-flight steering, altitude control, confidence signal, escalation triad |
| Verification and repair | How does anyone know the agent did what it says it did? | Artifacts as receipts, answer verification, action audit and time-limited undo, checkpoints, rationale grounded in precedent, designed repair |
| Reading a crowd of agents | What is this system of agents doing, and where is it stuck? | Orchestrator and workers, delegation briefs, goal hierarchy, conflict surfacing, mission control, agent identity |
| Many humans, many agents | How does a team share an agent without losing track of who did what? | Addressable team member, permission parity, trigger rights, external participants, agent registry |
| Ambient and asynchronous | How does work reach a person when there is no session? | Event triggers, agent inbox, scheduled agents, interruption budget, resumability |
| Agents as users | Can an agent operate a product built for people? | Machine-readable context, skills, the protocol stack, generative interfaces, delegated identity |
Where to spend the checking
Human attention is the scarce resource in agentic work, so oversight has to be placed rather than sprinkled. Three strategies are available and they perform differently.
Confirming only at the end is the current default in shipped products. It is also the weakest, because one error early in a long task means redoing everything after it. Confirming every step catches errors immediately and removes the reason to delegate at all.
Placing checkpoints deliberately sits between the two. Jieyu Zhou, Aryan Roy, Sneh Gupta, Daniel Weitekamp and Christopher MacLellan modelled this as a scheduling problem and tested it with 48 participants at CHI 2026, reporting that 81 percent preferred intermediate confirmation over confirming at the end, and that task completion time fell 13.5 percent. Their paper is arXiv:2510.05307.
Checkpoints earn most when placed early, because an error in step three is cheaper to catch than an error in step eleven is to unwind.
Eight failure modes
These are the failures that recur across products. Each one names the pattern that addresses it.
- 1Confirm only at the end
The agent works for twenty minutes and then asks whether it got the task right.
Addressed by scheduled confirmation.
- 2Transparency theatre
A wall of tool calls presented as reasoning. Overload reads as opacity, and nobody checks anything.
Addressed by altitude control.
- 3One autonomy switch for everything
A single setting covering every task the agent performs. One bad experience turns the whole feature off rather than dialling it back.
Addressed by an autonomy dial set per task type.
- 4Silent irreversible action
No preview before, no undo after, and no statement of when reversal stops being possible.
Addressed by intent preview and time-limited undo.
- 5Notification flood
Several background agents each interrupting on their own schedule. People stop reading, including the message that mattered.
Addressed by an interruption budget.
- 6Chat as the container for everything
Parallel, background and overnight work forced through one transcript.
Addressed by an agent inbox and a mission control view.
- 7Anonymous fleets
Agents with no name, no owner and no stated remit. Gartner named this pattern agent sprawl in April 2026.
Addressed by an agent registry.
- 8Confident guessing
Ambiguity resolved quietly into a plausible answer, which arrives finished and formatted.
Addressed by the escalation triad.
How this library is used in engagements
Discovery and direction work uses it as the checklist for an agent feature before it is built, so that handover and verification are specified alongside the capability.
Implementation and adoption work uses it as the diagnostic for a feature that shipped and is not being used, where the gap is usually handover or verification rather than model quality.
Related prior work
Shape of AI, by Emily Campbell, is a pattern library for AI product interfaces, published under Creative Commons BY-NC-SA.
Products cited on this page: Google Antigravity, Anthropic Claude Code and Cowork, Microsoft Magentic-UI, LangChain Agent Inbox, Asana AI Teammates, Notion Custom Agents. Research cited: Bansal and colleagues at Microsoft Research on human-agent communication, Zhou and colleagues at Georgia Tech on confirmation frequency, Schömbs and colleagues on multi-agent interaction, Lee on agent identity design.
Common questions
- What is an agent design pattern?
- A repeatable solution to a recurring problem in the interface between a person and an AI agent that takes actions. Examples include showing a plan before acting, placing confirmation checkpoints inside a long task, and producing reviewable deliverables the person can check in place of a log.
- How is this different from prompt engineering?
- Prompt engineering shapes what an agent is asked to do. These patterns shape how a person authorises, supervises and verifies the work. A well-prompted agent with no handover or verification design is still one people stop trusting after the first surprise.
- What is an intent preview?
- A statement of what the agent plans to do, in the vocabulary of the person who asked, shown before any consequential action, with at least three responses available: proceed, edit the plan, or take the task back. Victor Yocco named and specified the pattern in Smashing Magazine in February 2026, and argues it should be mandatory for actions that are irreversible, financial, or visible to a third party.
- What is an autonomy dial?
- A control that sets how much an agent may do without asking, on a four-rung scale: observe and suggest, plan and propose, act with confirmation, act autonomously. Setting it per task type rather than once for the whole product is what keeps a single bad experience from ending use of the feature.
- Should an AI agent ask for confirmation at every step?
- No. Confirming every step removes the benefit of delegating. Confirming only at the end is worse, because an early error forces a full redo. The tested alternative places a small number of checkpoints deliberately, weighted toward the start of the task.
Related: Prompting: the RIGHT framework · Discovery & Direction · Implementation & Adoption · AI Advisory: Strategy, Enablement and Adoption