AI agent
Yapay zekâ ajanı
D3
An AI agent is software that interacts with its environment, receives information, and undertakes self-directed actions in service of a larger, externally specified goal.
Review status: 2026-11-26
Technical explanation
An agent can plan or select actions, invoke tools, process their results, and feed those results into later model steps.
Conceptual boundaries
An AI agent is not defined by a single model response or tool call, and the cited sources do not establish independent human replacement or guaranteed autonomous competence.
Provider-neutral example
A research assistant can search an approved knowledge base, inspect returned passages, request another search when needed, and present a draft for human review.
Limitations
Agent definitions and implementations vary; tools, memory, and planning are possible features that can enlarge the security attack surface.
Related concepts
Atomic claims and evidence
1.1NIST AI 100-2e2025, Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations
- Source
- NIST AI 100-2e2025, Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations
- Source role
- Authoritative source
- Exact locator
- Appendix A, agent, printed p. 107
- Supported claim
- NIST defines an agent as software that interacts with its environment, receives information, and undertakes self-directed actions in service of a larger, externally specified goal.
- Last verification
- Review due
- Scope limitation
- This is NIST’s glossary definition and does not claim a single universal definition of all AI agents.
2.1Yao et al., ReAct: Synergizing Reasoning and Acting in Language Models
- Source
- Yao et al., ReAct: Synergizing Reasoning and Acting in Language Models
- Source role
- Authoritative source
- Exact locator
- Abstract: interleaved reasoning and task-specific actions; external sources or environments
- Supported claim
- In the ReAct research pattern, LLM reasoning traces and task-specific actions are interleaved, and actions interface with external knowledge sources or environments for additional information.
- Last verification
- Review due
- Scope limitation
- This describes one research pattern and does not require every agent to use ReAct, reasoning traces, or external tools.
3.1NIST AI 100-2e2025, Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations
- Source
- NIST AI 100-2e2025, Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations
- Source role
- Authoritative source
- Exact locator
- Section 3.5 printed p. 54: tools, memory, planning, and security risks
- Supported claim
- NIST states that agents may have tools, memory, or planning capabilities and that tool-enabled agents can create additional security risks.
- Last verification
- Review due
- Scope limitation
- These are possible features and risks, not required properties of every agent or proof of a particular agent’s autonomy.