# Monitor and Review

Once you publish an AI Agent and activate it, it starts doing real work. This section covers what happens from that point on: how a live flow processes incoming work, where items go for human judgment, and the views you use to follow runs and act on what needs attention.

This page gives you the operating model for production. The pages it links to cover each operational view in detail.

## How a published AI Agent runs

In production, a flow runs on its **trigger** — it isn't started by hand. The model is the same for every AI Agent:

1. **A configured trigger starts a run.** When the flow is active, its trigger brings in new input as it arrives — a manually uploaded file, an incoming email and its attachments, or another configured source — and starts a run.
2. **The flow's nodes process the input.** Each node reads the data available to it, does its work, and passes its outputs to the nodes connected after it. Classification, extraction, matching, routing, and system write-back all happen here, in the order you connected the nodes.
3. **A Review node can route work to a person.** Where you placed a **Flag for Human Review** node, the flow can hand the work to a reviewer for human judgment before continuing. The reviewer's decision returns the work to the flow.
4. **The flow completes or fails.** Based on the path you configured and the outcome of each node, the run finishes on its completion path or stops on a failure. Each piece of work the flow takes in moves through the run as an **item**, and its **status** reflects where it ended up.

Manual execution is a build-time tool, not the production model. **Run Draft** in the builder runs a draft against sample input so you can test it — see [Test and debug a draft](../build-ai-agents/test-debug-draft.md). In production, the trigger is what starts runs.

## The operational views

Three views give you the currently available picture of a live AI Agent:

- **Run history** shows the runs a flow has produced. Open a run to see how it executed step by step — its status, each node's execution details, the inputs and outputs, and any errors. Use it to confirm runs are succeeding and to investigate failures. See [Run history](./run-history.md).
- **Human review** is where items the flow flagged go for a person to act on. Reviewers inspect the source data and the information the flow produced, correct it where correction is available, and complete the review to return the item to the flow. See [Human review](./human-review.md).
- **Item status reporting** summarizes items by status so you can see, at a glance, how much work is in progress, waiting on review, complete, or failed. See [Item status reporting](./item-status-reporting.md).

## From building to operating

Monitoring picks up where building leaves off. You build, test, and publish an AI Agent in the [Build AI Agents](../build-ai-agents/overview.md) section; you **activate** it to put its trigger live, as described in [Drafts and publishing](../build-ai-agents/drafts-publishing.md). From then on, the views above are how you keep watch: confirm runs are healthy, clear the items waiting on review, and follow up on anything that failed. When you need to change the flow, you keep iterating in the draft and publish a new version — the active version keeps running in the meantime.

## Where to go next

- [Run history](./run-history.md) — browse and inspect a flow's runs.
- [Human review](./human-review.md) — act on items a flow sends for review.
- [Item status reporting](./item-status-reporting.md) — read item counts by status.
- [Drafts and publishing](../build-ai-agents/drafts-publishing.md) — publish and activate an AI Agent.
- [Configure triggers](../build-ai-agents/triggers.md) — how a live flow brings work in.