Open Flow
Browse documentation
DocumentationNodes
Beta 77Markdown version
On this page

AI Agent

Use a bounded model and tool loop with explicit data, tools, and approval policies.

Use AI Agent when the task needs several reasoning or tool-use rounds. Its declared model, tools, account choices, fixed arguments, and approval policies bound what the model may do.

Configure the task

Provide a prompt, select an available model, and bind the information the Agent needs.

{
  "op": "node.add", "as": "reply", "type": "agent", "name": "Draft a reply",
  "prompt": "Draft a brief reply. Identify facts that need confirmation. Do not invent a resolution.",
  "config": { "model": "deepseek-v4-flash", "maxRounds": 5 },
  "inputs": { "input": { "kind": "value", "value": "The customer's parcel has not arrived." } }
}

Select a model supported by the deployment. Named business inputs can be bound directly; use config.inputs when you need explicit constraints. Agent results default to text. Set config.resultSchema for a structured JSON result rather than declaring internal output ports.

An Agent receives a support request and produces a bounded reply draft. No external tools are granted in this example.

An Agent receives a support request and produces a bounded reply draft. No external tools are granted in this example.
Input
{"text":"My parcel has not arrived. Can you confirm its status?"}
Expected result
A reply draft that identifies facts requiring confirmation; no email is sent.
Before running
Configure the deployment's model service and select an available model before running.
Download workflow

Declare tools

Tools can be Connector Actions or selected published workflows. For each tool, configure its public name, description, account, and approval policy. Parameter sources can be fixed values, Agent inputs, or model-provided arguments.

The platform derives Action and workflow input definitions. The model cannot grant itself a new account, change fixed arguments, or disable approval. A deferred unready tool is excluded from the callable tool set.

Waiting and recovery

A tool can require human approval. The Run owns its pending decision and supported Agent continuation. Persisted tool results let recovery continue without repeating completed external calls to reconstruct data.

Temporary Agent code computation does not edit the Flow or automatically inherit external capabilities. Configure the supported code tool explicitly when required.

Limits

Agent has a bounded round count and node/Run deadlines. It does not support the ordinary node retry policy or onError handling used by Code and LLM. A model-generated result is not evidence that an external action succeeded; inspect actual tool results.

Sources

Search