Agentic AI,
under governance.

What agentic AI is

Agentic AI is software that pursues a goal rather than answering a prompt. It decomposes the goal into steps, decides which tools or sources each step needs, acts, reads what came back, and revises the plan when reality disagrees with it — stopping when the goal is met, or handing the decision to a person when it cannot be.

A generative system produces an output from an input. An agentic system pursues a goal: it decomposes the goal into steps, chooses which tools or sources each step needs, acts, observes what came back, and revises the plan when reality disagrees with it.

The loop is what distinguishes it. A model that answers once cannot recover from a bad assumption; a system that observes its own results can notice the totals do not reconcile and go back for the source document.

Planning, reasoning and tool use

Planning

Breaking a goal into an ordered set of steps, with the plan held as explicit state rather than improvised turn by turn.

Tool use

Calling databases, APIs and document stores through scoped, authenticated interfaces — each call narrow enough to be safe on its own.

Observation

Reading what the tool returned and testing it against what the plan expected, instead of assuming success.

Revision

Changing the plan when an observation contradicts it, and stopping when the goal is met or cannot be.

Escalation

Handing a decision to a person when confidence is low or the action crosses the approval threshold.

Multi-agent systems, and when not to use one

A multi-agent system is one in which several specialised agents divide a task between them and coordinate through defined handoffs, rather than one generalist agent carrying the whole thing.

Multiple specialised agents can outperform a single generalist when the subtasks genuinely differ — one agent that reads documents, another that queries structured data, a third that composes the result. Most deployments run a mixture-of-experts configuration for the same reason: several models behind one interface, each chosen for what it is good at.

They also multiply the failure modes. Every handoff is a place for context to be lost and for an error to be laundered into a confident downstream answer. A single well-scoped agent is very often the better engineering decision, and we will say so rather than building something more impressive than the problem requires.

Governance is the hard part

An agent that can act is an agent that can act wrongly. The controls that make autonomy acceptable in an enterprise — least privilege scoped to the session, time limits, a recorded trail of every action, an approval threshold above which a human decides — have to be properties of the runtime.

Written into a policy document instead, they are assumptions. Enforced in the system, they are constraints, and they hold whether or not the model behaves as expected on a given day.