Skip to main content
dagraph turns your AI workflows into plain YAML files that you version in git, review in PRs, and run from the CLI or trigger via webhook. Define a graph of agent calls, code execution steps, and human approval gates — dagraph runs parallel nodes automatically, retries on provider errors, and checkpoints every step so you can resume after a crash.

Quick Start

Run your first parallel agent workflow in under 5 minutes.

Node Types

Learn all 10 node types: agent, evaluator_loop, map, planner, and more.

CLI Reference

Full reference for every agentgraph command.

Multi-Provider

Configure fallback chains so workflows survive provider outages.

​
Why dagraph

Most AI workflow tools lock automations inside a UI. dagraph treats the YAML file as the source of truth — you own your workflows, can diff them in pull requests, and deploy them through your existing CI pipeline.

Parallel by default

Nodes with no dependencies run simultaneously. A 3-researcher + synthesizer DAG fires all three researchers at once.

Provider-agnostic

Claude, OpenAI, Gemini, Bedrock, and Ollama all work. Set a fallback chain per node and survive any outage.

Crash-safe

Every node status change is checkpointed to SQLite. Resume a paused or interrupted run without re-running completed nodes.

​
Get started

1

Install dagraph

pip install dagraph
2

Write a workflow

Create a YAML file that defines your agents and their dependencies.
3

Run it

agentgraph run my-workflow.yaml --input topic="quantum computing"
4

Inspect results

agentgraph inspect <run_id>
The default backend uses the claude CLI and your Claude Code plan — no API key required to get started.