Introduction

Welcome to the taskflow docs. taskflow is a declarative, verifiable graph of tasks for coding-agent subagents — you describe multi-step agent work as a DAG, the runtime verifies it before a single token is spent, and only the final result ever reaches your context window.

Here is the whole idea in four lines of JSON:

{
  "task": "Summarize the architecture of src/",
  "agent": "explorer"
}

That is a complete taskflow. The runtime turns it into a graph, runs one subagent, and returns the result. Add more phases, wire them with dependsOn, and the same shape scales to fan-out over hundreds of files — with retries, gates, loops, and cross-session resume handled for you.

Who is this for?

taskflow is for anyone who delegates multi-step work to a coding agent and has hit one of these:

• Context-window pressure. Every subagent transcript floods your conversation, and a five-step audit eats your context before you ever see a result.
• Plans that only exist while they run. You describe a workflow in prose, the model re-derives it every run, and there is nothing to diff, reuse, or hand to a reviewer.
• No recovery. A run dies at the last step and you start over from scratch — paying for every subagent call again.
• Reusable pipelines. You want to name a workflow once and invoke it by name, on Pi, Codex, Claude Code, or OpenCode.

If any of that sounds familiar, taskflow was built for you.

How to read these docs

The docs are organized as a learning path. Read them in order the first time; jump around once you know what you are looking for.

The chapters

Last updated on