agentenv/agentflow

Orchestrate thousands of agents and harnesses as a graph programatically

What it solves

AgentFlow is an orchestration framework that allows users to build complex, multi-agent workflows using dependency graphs. It solves the problem of managing multiple LLM agents (like Codex, Claude, and others) in a structured way, enabling parallel execution, iterative loops, and deployment across various isolated execution environments.

How it works

Users define a Graph where nodes represent agent tasks (e.g., using codex or claude functions). These nodes are connected via dependency edges (using the >> operator), defining the execution order. The framework supports:

  • Parallel Fanout: A single node's output can be dispatched to many parallel copies of another node using fanout(), which can be based on a count, a list of items, or a cartesian product.
  • Iterative Cycles: Nodes can be configured with on_failure triggers and success criteria to loop until a specific condition is met or a maximum number of iterations is reached.
  • Execution Targets: Agents can be run in isolated environments including local Docker containers, KVM VMs via Cloud Hypervisor, or remote machines (EC2, ECS Fargate, SSH).
  • Model Routing: Through the pi agent, the framework can route requests to a wide variety of of external providers (OpenAI, Anthropic, Groq, etc.) or local endpoints (Ollama, LMStudio).
  • Inference Setup: It integrates with SkyPilot to launch OpenAI-compatible endpoints on cloud GPUs.

Who it’s for

It is designed for developers and security researchers who need to coordinate multiple AI agents to perform complex tasks like automated code review, dependency auditing, and vulnerability discovery.

Highlights

  • Flexible Orchestration: Supports dependency graphs with parallel fanout and iterative loops.
  • Diverse Execution Targets: Provides high isolation through Docker, Cloud Hypervisor VMs, and cloud-native remote targets.
  • Cloud GPU Integration: Simplifies launching inference services via SkyPilot.
  • Tuned Agent Evolution: Allows creating reusable tuned agents based on successful Codex run traces.
  • Shared Memory: Includes a "scratchboard" for shared memory across all agents in a pipeline.
  • Web UI: Includes a local web interface and API for managing runs.

Related

  • Project
    google/adk-python

    A code-first Python framework for building, evaluating, and deploying sophisticated AI agents using graph-based workflows and modular multi-agent systems.

  • Project
    microsoft/agent-framework

    Microsoft Agent Framework is an open‑source SDK for building production‑ready AI agents and multi‑agent workflows in Python, .NET, and Go. It provides provider‑agnostic LLM access, graph‑based orchestration, middleware, observability, and easy Azure/Foundry hosting.

  • Project
    EpicStaff/EpicStaff

    EpicStaff is a self‑hosted, source‑available platform that lets teams design, run, and audit multi‑agent AI workflows via a visual node editor built on Django. It supports custom Python tools, LLM/RAG integration, persistent memory, webhook triggers, and enterprise security, giving operations teams ownership of AI processes while engineers handle system integration.

  • Project
    scaleapi/scale-agentex

    A framework for building and deploying AI agents that scales from simple chatbots to fully autonomous systems, supporting both synchronous and asynchronous workflows.