<span id="使用智能体创建工作流"></span>

# Create workflows with an agent

In the workflow editor, an agent helps you plan processing steps, build a candidate, and validate it through compilation checks and tests. Creating a candidate does not process your production data.

<span id="准备需求"></span>

## Prepare requirements

Open the workflow editor and choose the natural-language creation option. In the conversation on the right, specify the source data, required processing, models, and output location. Confirm that referenced resources are accessible in the current workspace.

For example:

> Read PDFs in the product manuals volume, extract product names, versions, and scope with supporting source text, and save the results to the specified output volume. Test one file first, then process the others after confirmation.

Inputs should distinguish source files from result locations. If no suitable location exists, create directories, databases, or volumes in the Catalog picker, then confirm the actually selected objects.

<span id="确认计划"></span>

## Confirm the plan

1. Review the proposed plan and its sources, steps, and outputs.
2. Explain any changes before confirming it.
3. Confirm the plan to begin building.

<span id="构建测试并应用"></span>

## Build, test, and apply

1. Inspect build progress and compilation results. Read the failure location before changing the plan or configuration.
2. Test with a small, verifiable input. Check field names, types, and syntax when supplying JSON.
3. Review the candidate canvas and bindings. Generated Python processing nodes are read-only; see [Python processing](nodes/code-processing/python-processing.md).
4. Inspect test outputs and destinations, then decide whether to apply the candidate.
5. After applying, check the saved workflow and follow [Run and debug](run-debug.md) to submit a production run.

Compilation, a successful test, deployment, and business completion are separate results. Inspect the run record and actual destination after running.

<span id="继续修改或恢复创建"></span>

## Continue editing or resume creation

After saving a workflow created through conversation, continue making changes in its associated agent conversation. Review the new candidate before applying it.

If creation was interrupted, use **Continue previous creation** when offered. Review the recovered goal, plan, and draft before continuing. Expand a collapsed conversation panel to inspect existing progress before submitting another request.

For failures, check resource access, required settings, bindings, test JSON, and the failing node. Preserve the workflow, conversation, time, and visible error. Confirm previous writes before repeating a test or run.

<span id="构建失败时"></span>

## Investigate build failures

| Symptom | Check and action |
| --- | --- |
| Cannot find the source or model | Confirm Workspace, resource location, and access privileges. Continue after changing the selection. |
| Compilation failed | Locate the failure and check required node fields, field bindings, and variable types. |
| Test input cannot be read | Check fields and value types in the test form or JSON. |
| Test run failed | Read the failed node's inputs, outputs, and errors before deciding which step to modify. |
| The result is unknown after interruption | Inspect drafts, associated conversations, and run history to confirm completed steps before continuing. |

When reporting a failure, retain the workflow, associated conversation, operation time, and page error message. For tests or production runs involving writes, confirm the outcome before submitting again.

<span id="程序化创建"></span>

## Create programmatically

For the API candidate flow, see [Workflows and templates](../../../../developer/api/ai-studio/data-processing/workflows-templates.md). For programmatic publishing, see the SDK guide [Confirm and Publish Workflow Proposals](../../../../developer/sdk/ai-studio/数据处理/确认并发布工作流方案.md) and the CLI guide [Workflows and Services](../../../../developer/cli/tasks/workflows.md#cli-workflow-candidates).
