Agentic AI
Agentic AI is a system that coordinates multiple AI agents, tools, and data sources to carry out multi-step work with limited human intervention.
Where a single AI agent typically handles one well-defined task, agentic AI describes the wider arrangement: several agents and tools orchestrated toward an outcome that spans systems or teams. In practice the two terms overlap heavily and vendors use them loosely.
The engineering difficulty is rarely the reasoning. It is bounding what the system can reach, requiring human approval before irreversible actions, and being able to reconstruct afterwards exactly which tools were called and with what arguments.
How it works
An agentic system usually has an orchestrator that breaks the goal into tasks and hands them to specialised agents, each with its own tools and instructions: one researches, one drafts, one checks. Results flow back to the orchestrator, which decides whether the outcome is good enough or another round is needed.
Shared state is the hard part. Agents need a common record of what has been done and decided, or they repeat work and contradict each other.
Example
A support workflow might route a customer email to one agent that classifies it, another that looks up the order in the store's database and a third that drafts a reply, with refunds above a set amount held for a person to approve before anything is sent.
Common mistakes
- Adding more agents to fix a problem that is really an unclear goal or a missing tool.
- Letting agents approve each other's irreversible actions with no human in the loop.
- Measuring the demo rather than the error rate across hundreds of real cases.
