What is an agent design pattern?

An agent design pattern is 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. Agent design patterns cover how a person authorizes, supervises and verifies the work.

What is an intent preview?

An intent preview is a statement of what the agent plans to do, shown before any consequential action and written in plain language. The person can respond in at least three ways: 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?

An autonomy dial is a control that sets how much an agent may do without asking. It has four levels: observe and suggest, plan and propose, act with confirmation, and act autonomously. It is set separately for each type of task.

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.

Pattern library

Agent design patterns

Agent design patterns describe how a person hands work to an AI agent, stays involved while it runs, checks what it did, and supervises several agents at once. We group them into seven families, and each pattern comes from a released product or a published research paper.

How an agent differs from an assistant

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 product teams handle that. They come from products released by Google, Anthropic, Microsoft, LangChain, Asana and Notion, and from human-computer interaction research published in 2025 and 2026. Each pattern refers to the product or paper it came from.

Emily Campbell’s Shape of AI is a pattern library for AI interfaces that covers the prompt-and-response loop in depth. This resource covers what changes when an agent works outside that loop:

Time: the work continues after the session ends, as with a research run, a migration or an overnight job, and 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 passes between the person and the agent. The person sets the brief. The agent drafts a plan, and the person edits it. The agent runs the task, and the person and the agent meet at checkpoints along the way. The person verifies the result. What the system learns feeds into the next brief.

Products built in 2025 and 2026 tend to be strong at the run stage and weak at the brief, verify and learn stages. This imbalance limits how much people delegate.

The seven families

HandoverWhat exactly was asked for, and what is the agent allowed to do?: Agreeing the plan and its limits before an agent starts costs less than undoing a wrong result afterwards. Many agent products built in 2025 and 2026 do some version of this.; Intent preview, Editable plan, Clarifying interview, Goal and task separation, Autonomy dial, Budget envelope, Durable briefs,
Oversight in motionHow does a person oversee an agent without watching every step?: Human attention is limited in agentic work. Oversight should be placed where an error would be expensive to catch late. Deciding where to place it is part of designing the product.; Scheduled confirmation, Action guard, Take over and hand back, Mid-flight steering, Altitude control, , Escalation triad
Verification and repairHow does anyone know the agent did what it says it did?: The product should make an agent’s work easy to check. Checking should not depend on how diligent the person is.; Artifacts as receipts, Answer verification, Action audit and time-limited undo, Rationale grounded in precedent,
Reading a crowd of agentsWhat is this system of agents doing, and where is it stuck?: Supervising each agent directly gets harder with every agent you add. Research on human interaction with robot swarms found that the effort needed to coordinate a group directly grows at least linearly with the size of the group. User studies find that people prefer orchestration once there is more than one agent to watch.; Orchestrator and workers, Delegation briefs, Goal hierarchy, Conflict surfacing, Mission control, Agent identity
Many humans, many agentsHow does a team share an agent without losing track of who did what?: Products are ahead of the research here. Asana, Notion and Slack shipped versions of the same idea within a year of each other, and the academic work on how groups make sense of work done by groups of agents is only now appearing.; Addressable team member, Permission parity, , Attribution in the shared record, Agent registry
Ambient and asynchronousHow does work reach a person when there is no session?: In a chat, the person has to start every exchange. That limits how much work one person can hand to agents. Agents that start work on their own when an event happens, such as an email arriving or a build failing, need a different kind of interface from chat.; Event triggers, Agent inbox, , Interruption budget,
Agents as usersCan an agent operate a product built for people?: Matt Biilmann of Netlify introduced the term agent experience in January 2025, describing the experience an agent has as a user of a product. It has four pillars: access, context, tools and orchestration.; Machine-readable context, Skills, The protocol stack, Generative interfaces, Delegated identity

Review an agent feature with us30 min

Where to place checkpoints

Human attention is limited in agentic work, so oversight has to be placed at chosen points in a task, using one of the three strategies described below.

Confirming only at the end is the current default in released products, and it is the weakest strategy. One error early in a long task means redoing everything after it. Confirming every step catches errors immediately, but it takes so much of the person’s attention that delegating brings little benefit.

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 are most useful early in a task. Catching an error in an early step costs less than undoing one that is found near the end.

Eight failure modes

These eight failures recur across agentic products.

1
Confirm only at the end
The agent works for twenty minutes and then asks whether it got the task right.
Addressed by scheduled confirmation.
2
Transparency theatre
The agent shows a long list of tool calls and presents it as its reasoning. There is too much to read and people stop checking the work.
Addressed by altitude control.
3
One autonomy switch for everything
One setting controls the agent’s autonomy for every task it performs. After one bad experience people will just turn the whole feature off.
Addressed by an autonomy dial set per task type.
4
Silent irreversible action
The agent acts without a preview first or without an undo option.
Addressed by intent preview and time-limited undo.
5
Notification flood
Several background agents each interrupt the person without a view of the bigger picture. People stop reading the messages and miss important ones.
Addressed by an interruption budget.
6
Chat as the container for everything
The product forces parallel, background and overnight work through one chat transcript.
Addressed by an agent inbox and a mission control view.
7
Anonymous fleets
Agents run with no name, no owner and no stated remit. Gartner named this pattern agent sprawl in April 2026.
Addressed by an agent registry.
8
Confident guessing
When a request is ambiguous, the agent picks a plausible reading without saying so. It then presents the result as a finished, formatted answer.
Addressed by the escalation triad.

The same eight, with what each one costs and the pattern that addresses it, are set out on the failure modes page.

Six of these patterns are also explained in short visual decks. You can download each one free, in English or Swedish.

Review an agent feature with us30 min

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 to diagnose an agent feature that has been released but is not being used. The cause is often in how work is handed over to the agent or how its results are checked.

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?: An agent design pattern is 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. Agent design patterns cover how a person authorizes, supervises and verifies the work.
What is an intent preview?: An intent preview is a statement of what the agent plans to do, shown before any consequential action and written in plain language. The person can respond in at least three ways: 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?: An autonomy dial is a control that sets how much an agent may do without asking. It has four levels: observe and suggest, plan and propose, act with confirmation, and act autonomously. It is set separately for each type of task.
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.