AgentFlow

Active
GitHub Python MIT

Description

Orchestrate thousands of agents and harnesses as a graph programatically.

Key Features

  • Dependency graph orchestration for codex, claude, and kimi agents with parallel fanout
  • Iterative cycles with stop conditions and automatic retry until success criteria met
  • Remote execution on SSH, EC2, and ECS Fargate with zero-config auto-discovery
  • Scratchboard shared memory for cross-agent state and context passing
  • Tuned agent evolution from completed runs for creating reusable optimized agents
  • Multi-model support via Pi routing to Anthropic, OpenAI, Groq, local Ollama, and more

Use Cases

💡 Multi-agent code review pipelines with parallel file analysis and merged findings
💡 Large-scale fuzzing campaigns with 128-shard fanout and batch merge
💡 Iterative code generation with review-fix cycles until quality gates pass
💡 Cross-model debate and cross-critique for robust solution validation
💡 Remote deployment workflows orchestrating agents across EC2 and ECS

Tags

orchestration multi-agent graph python

Categories

🧠 Multi-Agent 🔄 Workflow

Quick Start

Install: curl -fsSL https://raw.githubusercontent.com/shouc/agentflow/master/install.sh | bash. Or: pip install -e .[dev]. Quick start: from agentflow import Graph, codex, claude; with Graph('my-pipeline') as g: plan = codex(task_id='plan', prompt='Plan the work.'); impl = claude(task_id='impl', prompt='Implement: {{nodes.plan.output}}'); plan >> impl.
Visit GitHub

Related Projects