agentflow

Why we included this project

AgentFlow lets you treat a multi-step coding job as a graph instead of a single chat session. You define nodes for planning, implementation, and review, then connect them with simple operators; the tool can fan a node out to dozens of parallel copies and merge the results back. It works with the coding agents you probably already have, Codex, Claude Code, and Kimi, and can route through the Pi agent to reach external and local model endpoints. Beyond the graph, it handles the practical parts of running this at scale, from retry loops that keep going until a success condition to shared scratch memory between agents, and it can run nodes on remote SSH, EC2, or ECS machines. Teams doing agent-driven code reviews, batch refactoring, or heavy parallel generation will find the fanout and merge primitives directly useful, and the Python API keeps pipelines readable and versionable.