How does an AI agent work, step by step?

A task arrives. The platform loads the agent's brief and gathers the facts the task needs. The model proposes a step, the platform checks it against the agent's bounds, and if allowed a tool carries it out. The result goes back to the model, and the loop repeats until the task is done, handed off or held for a person.

What are the steps in one task?

Follow a single item through an operations exceptions queue at a bank: a payment that failed because the beneficiary's account number did not match the name.

1. The task arrives. The exception lands in a work queue with an id. The agent assigned to that queue picks it up.
2. The brief loads. The platform gives the agent its standing instructions: what it is for, what it may do, when to stop. The brief is fixed by the agent's role, not by the task.
3. Context is assembled. The platform fetches the payment details, the customer's account, any earlier contact about this payee, each stamped with its source.
4. The model proposes a step. It reads everything and proposes calling a tool: look up the payee in the customer's saved payees.
5. The platform checks. It evaluates the proposal against the agent's role, certificate, sources, the bank's limits and platform guardrails. Reading saved payees is within scope. Allowed, and filed.
6. The tool runs. The lookup returns a saved payee with a slightly different name.
7. The result feeds back. The model now proposes drafting a message to the customer asking them to confirm the payee's details.
8. Check and act again. Drafting is allowed at depth 2. The draft is created and held for a person to send, because this agent drafts but does not contact customers about payment failures directly.
9. The task ends or waits. The task records that it is waiting for the customer's reply, with a deadline.
10. Everything is filed. Each proposal, verdict and result sits in the record under the task id.

Which parts are the model and which are the platform?

Keeping that split clear explains most of what makes an agent safe or unsafe. A model that also holds credentials, keeps its own state and decides its own permissions is an agent with no outside check.

What happens when the task has many steps?

The loop runs as many times as needed. Each pass is checked on its own. A step allowed early on does not pre-approve later ones, which matters when the facts change mid-task. Reliability over long chains is a subject of its own, covered in whether AI agents can plan multi-step tasks reliably.

When one agent needs another's help, the handoff is another step: a task posted to the other agent, through the same checks. The formats involved are explained in how agent-to-agent protocols work.

What stays with a person?

In the example, sending the message to the customer stayed with a person, because of the agent's depth for that capability. In general, people keep actions above the agent's depth or the business's limits, actions that rest on unverified claims, and decisions with significant effects on individuals. How the hold-and-approve step should look is set out in how to design the hold-and-approve step for an AI agent.

How does this look across a whole queue?

At scale, the same loop runs across hundreds of items. The agent works the routine ones to completion within its depth. The rest arrive in a person's held queue with drafts ready. The team's time moves from gathering facts to deciding. That pattern is applied in how to work an operations exceptions queue with an AI agent. The layer that checks each step is described in the governed agentic operating system guide.