Agent
Goal-directed action with feedback
In general computing, an agent is a situated system with some control over its actions, responsiveness to its environment and goal-directed behavior. LLM-based agents are one implementation, not the whole category. [10] [14] Editorial interpretation
On this page
Definition comparison
Classical weak agency
Peer-reviewed; general computing
Wooldridge and Jennings characterize a weak notion through autonomy, social ability, reactivity and pro-activeness. This influential characterization is attributed, not presented as the only universal test. [10]
Institutional project scoping
NIST draft; identity and authorization
The NCCoE concept paper describes autonomous decisions and actions under limited human supervision toward goals. It is a draft project-scoping document, not a final general-agent standard. [12]
LLM engineering
First-party; architectural distinction
Anthropic distinguishes predefined workflows from agents whose LLM dynamically directs the process and tool use. Its usage is narrower than general agent theory. [13]
Scope
Begin with general software or hardware agency. For LLM engineering, explicitly identify the model, action interface and feedback loop; access to a tool alone does not settle agency. [10] [13] Editorial interpretation
Includes
- Editorial example: a goal-directed robot controller can be an agent without an LLM. A software agent can act through commands in a virtual environment. [10] [14] Editorial interpretation
- Bounded autonomy can coexist with approval gates. Record which decisions the system controls and which require a person. [12] [14] Editorial interpretation
Excludes and counterexamples
- A bare text-completion call does not, by itself, show goal-directed interaction with an environment under our test. [14] Editorial interpretation
- A fixed tool pipeline is not automatically an agent in the dynamic-control engineering sense, even if it contains an LLM and an API. [13] Editorial interpretation
Different usages and open boundaries
- Planning, long-term memory and self-learning are not universal prerequisites in our working definition. Capability claims should be tested separately from the name Agent. [10] [14] Editorial interpretation
- We retain autonomy, goal direction, environment feedback and action as Repository criteria. Social ability is explicit in the classical weak notion, but is not a mandatory gate in our narrower operational proposal. [10] [14] Editorial interpretation
LLM-based agents
An LLM-based agent places a language model within an action-and-observation system. ReAct interleaves reasoning with actions that obtain external information, allowing later reasoning to respond to outcomes. It is one architecture, not a requirement that every agent expose a chain of thought. [11] Editorial interpretation
Distinguish choosing an action, granting permission and executing it. A model’s proposed tool call does not itself execute a tool; the surrounding runtime provides that capability. [5] [14] Editorial interpretation
Operational criteria
Repository test: identify a goal, observations, an action interface, the decisions controlled by the system, and how outcomes affect later actions. For an LLM-based agent, trace a model-selected action, its execution and returned feedback. State stopping and approval conditions. [14] Editorial interpretation
Concept relationships
Only is-a means a strict subtype. Uses and supports/orchestrates describe composition and control.
A designed workflow may delegate an activity to an agent; this is composition, not a claim that every workflow is agentic. [14] Editorial interpretation
Harness supports/orchestrates Agent
When it supplies the running loop, tools and execution policy for an agent. [4] Editorial interpretation
Repository working definition
Repository · Proposed / not certified
Use Agent for a system with bounded autonomy, a goal, environment feedback and an ability to act. Use LLM-based agent when an LLM participates in selecting actions. The agent uses a model; it is not classified as an LLM subtype. [14]
No consensus decision recorded.
Sources and research gaps
Research notes
Institutional definition gap: the NIST material checked is a draft identity/authorization concept paper. No final, general-agent institutional definition was established in this review; the classical peer-reviewed account remains central.
- Intelligent Agents: Theory and Practice
Michael Wooldridge and Nicholas R. Jennings. Tier 2 / Peer-reviewed journal article; author-hosted text
The Knowledge Engineering Review 10(2), June 1995, 115-152; DOI 10.1017/S0269888900008122. What is an agent? A Weak Notion of Agency; stronger notions and other attributes. Accessed 2026-10-10.
Context: General software/hardware agency, predating LLM engineering.
Verification: Author HTML checked; its 1994 conversion timestamp is not the publication year. Journal issue and DOI verified at Cambridge.
- ReAct: Synergizing Reasoning and Acting in Language Models
Shunyu Yao and others. Tier 2 / Peer-reviewed conference paper; author manuscript
ICLR 2023 camera-ready manuscript; arXiv:2210.03629v3, 10 March 2023. Abstract and section 2; ICLR 2023 publication header. Accessed 2026-10-10.
Context: An LLM reasoning/action architecture; not a definition of every possible agent.
Verification: ICLR publication header and arXiv v3 camera-ready record checked 2026-10-10.
- Scaling Managed Agents: Decoupling the brain from the hands
Anthropic. Tier 3 / First-party engineering blog
8 April 2026. Session, harness and sandbox architecture. Accessed 2026-10-10.
Context: Architecture of a specific hosted agent product.
Verification: Original text and publication record checked 2026-10-10.
- Accelerating the Adoption of Software and AI Agent Identity and Authorization
Harold Booth, Bill Fisher, Ryan Galluzzo and Joshua Roberts; NIST NCCoE. Tier 3 / Institutional draft concept paper
DRAFT, February 2026; public comment 5 February-2 April 2026. Challenge Overview, printed page 5, lines 54-58. Accessed 2026-10-10.
Context: Identity and authorization project scoping. Official draft, not a final general-agent definition.
Verification: Original text and publication record checked 2026-10-10.
- Building effective agents
Anthropic. Tier 3 / First-party engineering blog
19 December 2024; live page includes subsequent tooling updates. What are agents?; Building blocks, workflows, and agents. Accessed 2026-10-10.
Context: Company-specific distinction between predefined LLM workflows and dynamic agents.
Verification: Original text and publication record checked 2026-10-10.
- Repository terminology guide
Repository internal proposal. Tier 4 / Repository internal proposal; not certified
Repository proposal 1.3 / not certified. Term-specific proposal sections and editorial policy. Accessed 2026-10-12.
Context: Internal proposals and editorial interpretations, not external consensus.
Verification: Versions 1.1 and 1.2 add the Schema proposal and the flow/governance experimental example while preserving prior proposal sections. All remain unratified; no adoption decision claimed. Version 1.3 changes branding and certification scope only; it does not certify any proposal.
Revision record
- 1.0 / 2026-10-10
Researched and rewritten with attributed definitions, typed relations and an unratified Repository proposal.
- 1.1 / 2026-10-12
Scope the working definition and certification labels to this repository under Open AI Glossary. Citations and proposed status remain unchanged.