Skip to main content

Agent Executor

Category: Agent
Node id: agent-executor

Orchestrates an LLM agent with iterative tool-use capabilities

What it does​

  • Iterative LLM reasoning with tool-use loop
  • Configurable maximum iteration count
  • Tool call detection and delegation to pipeline orchestrator
  • Full message history tracking across iterations
  • System prompt customization for agent persona
  • Automatic stop detection on completion

Use cases​

  • Building ReAct-style agents that reason and act iteratively
  • Autonomous task execution with tool augmentation
  • Multi-step clinical workflows requiring external tool calls
  • Complex question answering with intermediate reasoning steps

Runtime​

  • Performance tier: slow

Input ports​

NameTypeRequiredDescription
tasktextYesThe task for the agent to execute
contexttextNoAdditional context for the agent

Output ports​

NameTypeRequiredDescription
resultjsonNoExecution result with answer, steps, iterations, status, and optionally tool_calls and messages
successbooleanNoWhether the execution completed without error

Settings​

Basic​

SettingNameTypeRequiredDefaultAllowed valuesDescription
LLM Endpointllm_endpointstringYesBase URL of the LLM API
ModelmodelstringNogpt-4oModel name to use
System Promptsystem_prompttextNoYou are a helpful AI assistant.System prompt for the agent

Auth​

SettingNameTypeRequiredDefaultAllowed valuesDescription
API Keyapi_keysecretNoAPI key for the LLM provider

Advanced​

SettingNameTypeRequiredDefaultAllowed valuesDescription
Max Iterationsmax_iterationsnumberNo10min 1; max 100Maximum number of reasoning iterations
TemperaturetemperaturenumberNo0.7min 0; max 2Sampling temperature
Max Tokensmax_tokensnumberNo4096Maximum tokens per LLM call
ToolstoolsjsonNoTool definitions for the agent to use

Errors​

CodeDescription
ENDPOINT_REQUIREDllm_endpoint is not configured
TASK_REQUIREDNo task input was provided
LLM_REQUEST_FAILEDLLM API call failed during an iteration
HTTP_ERRORLLM endpoint returned an HTTP error status

Runtime notes​

  • When tool_calls are pending, status is 'tool_calls_pending' and the messages array contains the full conversation history for resumption
  • result.steps is an ordered array — each entry records one LLM round-trip with its finish_reason
  • The node stops iterating on finish_reason 'stop' or when max_iterations is reached
  • Wire the success output to a branch node to handle LLM errors vs tool-call delegation

See Build any workflow for how to put nodes together and Ports, types and data flow for how port types connect.

Something unclear? Send documentation feedback by email.