agentflow

Why we included this project

AgentFlow suits teams that want to run many coding agents as a single coordinated job. You describe the workflow as a graph in Python, so a planning step can fan out into dozens or hundreds of parallel workers whose outputs merge in batches or groups before the next stage begins. It drives Codex, Claude, Kimi, and Pi agents, and the same graph will run on a laptop, through Docker, or across SSH, EC2, and ECS targets, which matters when you want to iterate locally before scaling out. The engine also supports iterative cycles with stop conditions and retry handling for runs that need several passes before they converge. A bundled Codex skill accepts a natural-language description of the task and writes and runs the graph, so teams already working inside a coding assistant can skip the tooling.