Glossary · term

Agentic AI

Agentic AI refers to AI systems that pursue a user-defined goal through a sequence of decisions and actions with some meaningful runtime autonomy. A system may plan, select and use tools, access data, observe results, revise its approach, and decide what step to take next instead of producing one response only. It may contain one agent or several cooperating agents. There is no universal threshold for how much autonomy makes a system agentic, so the label should be accompanied by a concrete description of its actions, permissions, and human checkpoints.

Agents2023-12-14Wave 1 · 2023Maturity: 3/5

Origin and context

Software-agent research predates the recent generative-AI cycle by decades. OpenAI's December 2023 governance paper supplies the earliest direct use of the exact label verified for this review and defines agentic AI systems around pursuing complex goals with limited direct supervision; it does not establish coinage. The Associated Press traces the label's wider industry prominence to 2024 and documents vendor-dependent usage. By 2026, Singapore's IMDA had published a governance framework and the United States' NIST had launched an AI-agent standards initiative, while IMDA still noted the absence of a universally accepted definition.

Sources: s4, s1, s2, s3

Why it matters

The shift from answering to acting changes both utility and risk. A system that can browse, write code, update records, contact services, or delegate work can complete longer tasks, but errors can propagate across steps and affect external systems. Evaluation therefore needs to cover trajectories, tool use, permissions, resource limits, recovery, and outcomes rather than only the quality of a final message. Human accountability remains in place even when the system chooses intermediate steps independently.

Sources: s4, s1, s2, s3

Example

An agentic procurement assistant might turn a request into a plan, search approved catalogs, compare offers, ask a supplier API for availability, and prepare a purchase order. Its autonomy should be stated precisely: it may read approved data and draft an order, while a person must authorize the transaction. Logs should capture tool calls and changes, credentials should be scoped to the minimum required access, and the system should stop or escalate when evidence is insufficient, costs exceed a bound, or the requested action falls outside policy.

Sources: s1, s2

How it differs

Agentic Workflows

Agentic AI is the umbrella system description. An agentic workflow is the multi-step process or orchestration pattern through which model calls, tools, and feedback are coordinated. Some taxonomies reserve workflow for predefined paths and agent for dynamic control; other sources use agentic workflow more broadly, so the control boundary must be stated.

Agentic Coding

Agentic coding is a domain-specific application of agentic AI to repository-level software work. Agentic AI also covers research, operations, customer service, and other tasks. Editing files or running tests can demonstrate action capability, but it does not define the full category.

Maturity and evidence

Maturity is rated 3. The term is used by independent public institutions and is tied to a stable operational core: goals, multistep decisions, tools, actions, and variable autonomy. It is not rated higher because boundaries remain contested, vendor marketing often stretches the label, and measurement and governance practices are still developing.

Sources: s4, s1, s2, s3

Limits and open questions

Agentic does not mean fully autonomous, generally intelligent, continuously learning, or reliable. A scripted pipeline with fixed branches may be marketed as agentic, while a genuinely dynamic system may still have a narrow action space. Claims should specify what the system can observe, decide, change, and delegate; which tools and data it can reach; how long it can run; and where approval is required. Because plans and tool results can fail, deployments need least privilege, sandboxing where appropriate, bounded resources, monitoring, evaluation on realistic trajectories, and recovery procedures. The label alone is not a safety or performance claim.

Sources: s1, s2, s3

Related terms

References

Last updated: 2026-09-04

In the Skills Atlas

This term is also covered in the Skills Atlas as ai agent design skill.

In the Skills Atlas

This term is also covered in the Skills Atlas as agentic planning task decomposition skill.

In the Skills Atlas

This term is also covered in the Skills Atlas as agent sandboxing skill.