What Is Agentic AI?

Agentic AI, also written as agentic artificial intelligence, refers to systems built around large language models that pursue multi-step goals with limited human intervention by planning a sequence of actions, invoking external tools, retaining state across those steps, and revising the plan in response to what the tools return. The distinction from conventional generative AI is behavioral rather than architectural: a chat model produces text in a single turn, while an agentic system runs a control loop in which model output becomes an action, the action changes some external state, and the result feeds back as new input. Systems of this kind combine autonomous planning, tool use, and memory over long-horizon interactions, with or without human oversight at intermediate steps.

The idea inherits from decades of work on autonomous agents in symbolic AI and reinforcement learning, but the current form dates from 2023, when reliable function calling and longer context windows made it practical to let a language model drive the loop directly. The U.S. National Institute of Standards and Technology maintains a program on agentic AI covering trustworthiness, evaluation, interoperability, and risk management.

Planning, Tool Use, and Memory

Three capabilities define the category. Planning decomposes a stated goal into ordered subtasks, often by interleaving explicit reasoning steps with actions so that each observation can revise the remaining plan. Tool use connects the model to code interpreters, search indexes, databases, browsers, and application programming interfaces through structured call formats, and interoperability standards have emerged to describe those connections uniformly rather than per vendor. Memory separates the working context of the current episode from durable stores that persist across sessions, typically retrieved by embedding similarity and injected selectively to stay within context limits. A survey of agentic tool use in large language models sorts the published methods into prompt-level plug-and-play interfaces, supervised tool learning, and reward-driven tool policy learning, and reviews how each is evaluated and where each fails.

Multi-Agent Organization

Many deployments distribute work across several specialized agents rather than one general one. A common pattern assigns an orchestrator to decompose the task and dispatch subtasks to workers with narrower instructions and tool permissions, then aggregates their outputs. Other configurations use debate among peer agents, hierarchical review, or a critic that evaluates a generator's work before it is accepted. The advantages are context isolation and parallelism. The costs are coordination overhead, compounding error across handoffs, and a much larger surface for failures that no single agent can observe. Analysis of trust, risk, and security management in multi-agent systems treats these organizational choices as governance decisions rather than purely engineering ones.

Evaluation and Security

Measuring agentic systems is harder than scoring single-turn outputs, because a task can fail at any of dozens of steps and partial progress resists simple scoring. NIST has approached this through work on building evaluation probes into agentic AI, which instruments the loop to observe intermediate decisions rather than only the final answer. Security concerns are distinct from those of static models: instructions embedded in retrieved documents or tool output can hijack the control loop, granted permissions can be exercised in unintended combinations, and a persistent memory store becomes an attack surface that survives the session. Published work on agentic AI security threats and defenses groups mitigations into sandboxing, least-privilege credentialing, action confirmation checkpoints, and runtime monitoring.

Applications

Agentic AI has applications in a range of fields, including:

• Software engineering, including code generation, testing, and repository maintenance
• Customer service and IT operations workflow automation
• Scientific literature review and experiment planning
• Financial and legal document analysis with tool-verified retrieval
• Robotics and embodied systems that use language models for task-level planning