An agentic AI system uses a language model not just to generate a single response, but to plan and execute a sequence of actions — calling tools, inspecting results, and deciding what to do next — toward a defined goal, typically with limited human supervision during execution.
What distinguishes "agentic" from "a chatbot with tools"
The distinguishing feature is the control loop: the model's own output determines what happens next, not just what text is displayed. A system that calls one tool and returns the result is tool-augmented; a system that calls a tool, evaluates whether the result actually accomplished the goal, and decides whether to retry, branch, or call a different tool next is agentic.
Common orchestration patterns
- Graph-based orchestration (e.g. LangGraph) — the workflow is an explicit graph of states and transitions, with the model's output determining which edge to follow. This makes control flow inspectable and supports retries, branches, and checkpoints.
- Role-based orchestration (e.g. CrewAI) — multiple agents are defined by role and goal, and a process coordinates handoffs between them, mirroring how a small team divides work.
Where these systems fail in practice
The gap between an agent that works in a demo and one that works unattended in production is almost always about what happens around the model's decisions, not the decisions themselves — validating tool-call arguments before executing them, making side-effecting steps safe to retry, enforcing cost ceilings, and giving the system a legitimate "escalate to a human" exit rather than letting it guess under low confidence.