-
Notifications
You must be signed in to change notification settings - Fork 0
Quickstart
From zero to a running multi-agent graph in about five minutes.
Grab the latest release for your platform from the releases page (Linux / macOS / Windows, SHA256SUMS included). Builds from main are official; builds from dev are prereleases.
Or build from source with Bun 1.3+:
git clone https://github.com/LeXwDeX/OpenCode-GraphAgent.git
cd OpenCode-GraphAgent
bun install
bun devCreate ~/.config/opencode/dag.jsonc (project-level .opencode/dag.jsonc also works and wins):
advanced runs the judgment nodes (reviews, arbitration, anything marked required). standard runs the volume work (exploration, implementation). One tier is fine to start — it becomes the default for everything.
In the TUI, type:
/dag-flow audit this repository's dependencies and propose an upgrade order
The agent designs a graph (exploration → analysis → report), starts it, and tells you the workflow ID. You do not need to define anything for this to work.
Open the command palette (ctrl+k), run dag.open. You get the workflow list, the node view in wave order with live status, and node detail on selection. Keys: p pause, r resume, s step, x cancel, enter drops into a node's child session.
Or from a script, the same information is on the HTTP API: GET /dag/<workflowID>/nodes.
Within a couple of minutes: waves of nodes flipping from queued to running to completed, the report node waking the parent with its output, and the workflow reaching completed. If a node fails (quota, API error), it shows as failed on the graph — fix or let the agent replan it, and the superseded segment disappears from the view while the workflow finishes on the replacement.
- Concepts — the vocabulary behind what you just saw.
- Authoring Workflows — save this procedure as a reusable, named workflow.
{ "model": { "advanced": "anthropic/claude-sonnet-4-5", "standard": "anthropic/claude-haiku-4-5" } }