AI Agent Executor node¶
| Node type ID | flowdrop_agents_ai_agents_executor |
| Executor plugin | flowdrop_agents:ai_agents_executor |
| Category | Agents |
Inputs¶
agent_id (required)¶
The agent to run. Pick it in the node form; the field cannot be connected. Validation fails if the ID is empty or no agent with that ID exists.
task_description (required)¶
The task for the agent. Connect it from an upstream node. It becomes the last user message sent
to the agent.
task_title¶
Optional short title, set on the agent's task object.
chat_history¶
Earlier messages, connected from an upstream node. An array of objects with role and content:
[
{"role": "user", "content": "Create a new article"},
{"role": "assistant", "content": "What should the title be?"},
{"role": "user", "content": "Use 'AI in Drupal' as the title"}
]
Every item must have both keys or validation fails. Only the roles user, assistant and tool
are passed to the agent. Messages with any other role, including system, are dropped silently.
Chat history often carries data the workflow author does not control, such as a visitor's chat
transcript, and a system message there could rewrite the agent's instructions.
create_directly¶
Boolean, default true.
true: the agent writes configuration and content as it goes.false: the agent returns a blueprint and changes nothing.
additional_context¶
Extra instructions. Set it in the form or connect it. It is appended to the task as
Additional context: ... in the same user message.
llm_model¶
Overrides the AI provider and model for this run. Connect a string in the AI module's combined
provider__model format, for example openai__gpt-4o. An empty string keeps the agent's own
provider and model. An unknown provider is ignored.
tool_usage_limits¶
Overrides tool parameter constraints for this run, keyed by tool ID and then by property name. Applies to config-based agents only; plugin-based agents ignore it.
The node merges its limits over the agent's own: per tool, then per property. A property the node
sets replaces the agent's setting for that property; every other limit the agent defines, including
force_value and hide_property guardrails, stays in place.
The usual source is the RAG Search Config node from FlowDrop AI Search, which builds these limits for a search index.
Outputs¶
| Output | Type | Exposed | Set when |
|---|---|---|---|
response |
string | yes | Always. The agent's answer, or a default message. |
agent_id |
string | yes | Always. |
agent_type |
string | yes | Always. Plugin or Config. |
job_status |
integer | yes | Always. See Job status. |
job_status_label |
string | yes | Always. |
success |
boolean | yes | Always. false only when the task is not solvable. |
structured_output |
object | yes | Solvable runs. |
error_message |
string | yes | Not solvable runs. |
information |
string | no | Informs runs. |
questions |
array | yes | Declared, but not populated. Questions reach a person through an interrupt instead; see Job status. |
execution_metadata |
object | no | Always. |
structured_output¶
Six lists, as returned by the agent's structured output:
{
"created_configs": [],
"edited_configs": [],
"deleted_configs": [],
"created_contents": [],
"edited_contents": [],
"deleted_contents": []
}
The shape of each entry comes from the AI Agents module.
execution_metadata¶
Always contains:
timestamp: Unix time of the run.create_directly: the value used.resumed_from_interrupt:truewhen the run continued after a person answered a question.
On solvable runs it also carries a debug trace:
- Config-based agents:
tools_executed(each withfunction_name,tool_idandoutput) andllm_conversation(the agent's chat history as arrays). - Plugin-based agents that implement
getExecutionTrace():tools_executedwith that trace.
The trace can contain content the agent read or wrote. Keep that in mind before you send
execution_metadata to a log or an external system.