# Agent Workflows

> Build, test, publish, run, and inspect multi-step operational workflows.

Canonical URL: https://marcenta.ai/docs/agent-workflows
Markdown URL: https://marcenta.ai/site-md/docs/agent-workflows

Workflows connect triggers, data, decisions, AI-assisted analysis, human review, and actions on a visual canvas. Use them for repeatable processes that would otherwise require several manual handoffs.

## Availability and prerequisites [#availability-and-prerequisites]

Workflow access depends on the plan and features enabled for your workspace. A workflow can use only data sources and destination connections already available to that workspace.

Before building, identify the trigger, required input data, decision points, person responsible for approvals, and intended final action. For any external write or email, confirm the destination and audience first.

## Create a workflow [#create-a-workflow]

1. Open **Workflows** and select **New workflow**.
2. Give the workflow a clear operational name.
3. Add one trigger.
4. Add downstream nodes and connect them in execution order.
5. Select each node and configure its inputs.
6. Bind values from earlier node outputs where required.
7. Test the draft and resolve validation errors.
8. Publish the validated workflow.

Draft changes save automatically. Wait for the saved state before leaving the editor or publishing.

## Available node types [#available-node-types]

The current editor can expose the following nodes, subject to workspace access and valid connections:

| Category | Nodes                                                                     | Purpose                                                                  |
| -------- | ------------------------------------------------------------------------- | ------------------------------------------------------------------------ |
| Trigger  | Manual Trigger, Schedule Trigger                                          | Start on demand or according to a configured schedule                    |
| Logic    | Condition                                                                 | Send a run down different branches based on a comparison                 |
| AI       | Custom Agent, Data Analysis Agent, Query Agent, Search Terms Review Agent | Generate or analyze structured output from supplied or connected context |
| Human    | Human Approval                                                            | Pause a run for an explicit review decision                              |
| Data     | Run Saved Report                                                          | Execute an existing report and provide its result to later steps         |
| Action   | Create Artifact, Create Task, Email Share, Push to Google Ads             | Persist, assign, deliver, or apply an approved result                    |

Node availability can change as capabilities are enabled. The node picker in your workspace is the authoritative list.

## Configure inputs and bindings [#configure-inputs-and-bindings]

A node can use fixed values, outputs from an earlier connected node, or both. When selecting a variable, verify that its type and meaning match the destination field. For example, a task title expects text, while a destination action may require a specific artifact and review reference.

Changing or deleting an upstream node can invalidate later bindings. Reopen downstream nodes after reorganizing the graph and resolve every validation message before publishing.

## Conditions and branches [#conditions-and-branches]

A Condition evaluates a comparison and directs the run along the matching branch. Keep conditions explicit and test values on both sides of the branch. A workflow that succeeds only on the expected branch has not fully tested its decision logic.

## Human approval and Google Ads changes [#human-approval-and-google-ads-changes]

**Push to Google Ads** is an external write action. It must follow a single connected **Human Approval** step and use the approved artifact and review produced by that step. This requirement prevents an analysis path from bypassing the review gate.

During approval, verify the target account, proposed changes, supporting data, and scope. Approval authorizes the configured downstream action; it should not be treated as a general endorsement of unrelated changes.

## Test, publish, and run [#test-publish-and-run]

Use a test run to validate configuration, branching, data bindings, and expected outputs. Node-level status and validation messages identify where a draft needs attention.

Publishing creates the validated version used for production runs. Publishing can be blocked when required configuration is absent, bindings are invalid, or the latest draft has not saved.

After publishing:

* use **Run now** for an on-demand run; or
* allow a Schedule Trigger to start runs on its configured cadence.

Changing a draft does not retroactively change a run that already started.

## Run history [#run-history]

Open **Runs** in the workflow editor to inspect previous executions. Review the overall run status and each node's progress or error. When a run fails, identify the first failing node; later nodes may be skipped because their required input never existed.

## Safe operating checklist [#safe-operating-checklist]

* Test every meaningful branch.
* Use a narrowly scoped destination connection.
* Require a human review before external writes.
* Confirm email recipients and data sensitivity.
* Give created tasks and artifacts recognizable names.
* Inspect the first production runs before relying on a schedule.
* Disable or revise a workflow when its source report, provider account, or business process changes.

<Callout type="warn">
  A test validates the workflow path; it is not a substitute for human approval of production-impacting actions.
</Callout>