AI agent patterns: tools, state and control

Separate planning, action and verification

A plan proposes steps; the executor checks permissions and arguments; observation records what happened. Keeping those functions distinct lets you change a tool without rewriting the entire flow. Save the state needed to continue and define when it expires. Memory should serve the task rather than accumulate data without a clear purpose.

A concrete example

An agent preparing a report might search documents, extract figures and draft text. Before sharing, it must check recipients, sources and permissions. If a search fails, it can record the gap and continue independent work; it should not invent evidence to finish the plan.

What to check

Test tools with errors, conflicting responses and uncertain operation outcomes. Limit attempts, cost and external actions. Evaluate deliverables and traces: a successful chain of calls does not prove the final objective was achieved.

Agent patterns

More autonomy requires better boundaries and observability.

PLAN

Plan and execute

Separates intention, tool selection and verification.

Control MEMORY

Task status

Keep only the necessary information and define its expiration.

Context REVISION

Humans in the circuit

Stop sensitive actions before producing external effects.

Supervision

ANALYSIS

Related analysis

GUIA

Persistent Memory for Agents: What to Store, for How Long, and How to Correct It

A practical design guide to memory records, lifecycle policies, and verification tests for AI agents—distinguishing persistent memory from conversation history and temporary task state.

16 Oct 2026 GUIA

Did the Agent Really Finish? How to Verify the Result Without Trusting Its Final Answer

An action being issued, a tool responding, and a task being completed are different things. Learn to define an observable result, choose appropriate evidence, and communicate precisely what has been verified.

16 Oct 2026 GUIA

Fixed Workflow, Single Agent, or Multi-Agent System: How to Choose a Pattern for a Task

A guide to comparing three ways to automate multi-step tasks based on variability, control, verifiability, and coordination costs. It includes a shared example and questions to help justify a choice—even when the best option is not to use an agent.

14 Oct 2026 GUIA

Permissions for AI Agents: How to Limit Tools, Data, and Actions to What’s Necessary

An operational guide to designing and testing agent permissions by task, user, resource, and duration. Includes an action matrix, human-approval criteria, and revocation checks.

27 Sep 2026 GUIA

Failure Recovery for AI Agents: When to Resume, Undo, or Stop

A practical protocol for recovering tasks with external side effects without mistaking a failed call for an action that never happened.

24 Sep 2026 GUIA

AI Agent Memory: Separating Context, Preferences, and Persistent Facts—and Deciding When They Expire

Useful memory is not an unlimited archive of conversations. It is a governed set of data with an owner, purpose, provenance, confidence, scope of use, and rules for review, correction, and deletion.

22 Sep 2026 GUIA

Humans in the Loop for AI Agents: How to Design Approvals That Reduce Risk Without Blocking Work

A useful approval policy does not merely add an “approve” button. It must state which actions are stopped, what evidence the reviewer receives, who can decide, and what happens when there is no response. This guide provides an operational method for turning an agent workflow into verifiable controls proportionate to risk.

22 Sep 2026 GUIA

Tool-Using Agents: How to Design Permissions, Memory, and Failure Recovery Without Giving Up Control

A guide to deciding how much autonomy to grant an agent that queries systems or performs actions. It explains how to define tools, permissions, memory, human approvals, recovery, and auditing before moving into production.

19 Sep 2026

Keep exploring

SourcesNIST AI RMFwww.nist.gov