Agentic AI

Agentic AI is the umbrella term for AI systems that pursue a goal over several steps, deciding what to do, using tools and memory, and acting on the results with some degree of autonomy.

A spectrum. Anthropic uses “agentic systems” for both fixed workflows, where code sets the path, and agents, where the model sets it. Between the two sit common patterns:

• Prompt chaining: each model call works on the previous one’s output.
• Routing: classify the input and send it to a specialized path.
• Parallelization: split work into sections, or run the same task several times and vote.
• Orchestrator-workers: a lead model decides the subtasks and delegates them.
• Evaluator-optimizer: one call drafts, another critiques, in a loop.

Most production systems sit somewhere on this range, and autonomy is a design choice made step by step.

What changes with autonomy. Once a model can act, its mistakes become actions. The OWASP Top 10 for Agentic Applications, published in December 2025, lists risks that a chatbot that only answers does not face, including agent goal hijack, tool misuse, identity and privilege abuse, memory and context poisoning, insecure inter-agent communication, cascading failures and rogue agents. The usual controls are narrow tool permissions, guardrails, human approval for consequential steps, and traces of what the agent did.

Neutral building blocks. Key agentic standards now sit in neutral foundations. The Linux Foundation launched the Agent2Agent (A2A) project in June 2025. In December 2025, Anthropic donated the Model Context Protocol to the Agentic AI Foundation, a directed fund under the Linux Foundation.

Neighbouring terms. An AI agent is the individual system. A multi-agent system coordinates several of them. Human-in-the-loop and guardrails are the two main ways to bound what they do.