Architecture

Flowgentic sits between an application's agent code and RADICAL AsyncFlow. It has two independent parts: bind_agent, which runs an agent component as an AsyncFlow task, and LangraphIntegration, which runs a LangGraph graph's tools, services, tasks and nodes as AsyncFlow tasks. AsyncFlow schedules every task on a named backend. A third part, accept, decides whether an agent's output may become that work; it calls no AsyncFlow API itself, though its producer and checks may submit tasks.

flowchart LR
  subgraph Application
    APP[AsyncFlow application]
    GRAPH[LangGraph graph]
  end
  subgraph Flowgentic
    BIND[bind_agent]
    LGI[LangraphIntegration<br/>asyncflow decorator]
  end
  APP -- agent component --> BIND
  GRAPH -- tools, services, tasks, nodes --> LGI
  BIND --> ENGINE[AsyncFlow WorkflowEngine]
  LGI --> ENGINE
  ENGINE --> BACKENDS[Named backends<br/>ai, compute]

Modules

Module	Contents
flowgentic.agent	bind_agent and AgentTask: an agent component as an AsyncFlow task, each call an invoke_agent span; see Telemetry
flowgentic.acceptance	accept, Rejected, NotAccepted, Outcome and Attempt: an agent's output checked, re-asked for and accepted or replaced by a fallback, each call an accept span; see Acceptance
flowgentic.langGraph.main	LangraphIntegration: creates and shuts down the WorkflowEngine
flowgentic.langGraph.execution_wrappers	AsyncFlowType, the asyncflow decorator, the handoff tool, BaseLLMAgentState
flowgentic.langGraph.utils	An LLM router and fan-out for supervisor graphs
flowgentic.langGraph.base_components	BaseToolRegistry: a base class that registers an application's tools and tasks by name
flowgentic.utils.llm_providers	ChatLLMProvider: OpenRouter, OpenAI and Ollama chat models
flowgentic	configure_logging: prints Flowgentic's log records, which import leaves to the application; see Logging
flowgentic._asyncflow	Private: the one module that calls AsyncFlow. bind_agent, LangraphIntegration, the asyncflow decorator and the handoff tool open and close engines and register tasks, services and blocks through it. It carries the caller's trace context into each task and service
flowgentic._telemetry	Private: the OpenTelemetry tracer and logger, the span, event and attribute names, the content switch, and the trace-context carrier

Academy agents' actions run in AsyncFlow tasks without a Flowgentic module: see Academy actions as AsyncFlow tasks.

The API reference documents each module.

How a decorated call runs

Each asyncflow decorator registers the function with AsyncFlow. A call submits one task or block and awaits it, with no retry and no timeout:

flowchart LR
  CALL[Graph or node code calls<br/>the decorated function] --> SUBMIT[AsyncFlow task or block<br/>on its backend]
  SUBMIT --> RESULT[Result, or the service's future]
  RESULT --> CALL

A node is usually an EXECUTION_BLOCK, and the tools, tasks and services it calls are AsyncFlow tasks inside that block. AsyncFlow's documentation describes blocks.

flowchart TD
  BLOCK[EXECUTION_BLOCK<br/>a graph node as an AsyncFlow block]
  BLOCK --> SERVICE[SERVICE_TASK<br/>resident service]
  BLOCK --> TOOL[AGENT_TOOL_AS_FUNCTION<br/>LangChain tool]
  BLOCK --> TASK[FUNCTION_TASK<br/>task]

LangGraph on AsyncFlow with Flowgentic describes each flow type, and Retries and timeouts where retries and deadlines belong.