A flow you defined
Predefined steps, dependencies and input mappings. Shows its definition version and real nodes with their inputs and outputs. It can contain zero AI.
DAG Flow · Visible agent control
Agents you can see, steer and keep.
DAG Flow is how Matrix runs work. Every Task becomes a graph of people, agents, groups, tools and explicit iterations, drawn on one canvas that shows the plan, the live run and the history together. You decide what runs next, and you keep the flows that worked.
Plan, run and history on one canvasStop, continue, retry, changeReusable flows and Skills
Task run · DAG Flowlive run view
Done, with evidenceRunning nowWaiting or not startedHand: real files and tests
What the graph shows.
Collaboration edges carry messages, delegation and results. Dependency edges say when a step is ready. Data edges show where inputs come from. Iteration edges feed the next round. They are never mixed into one topological sort.
HumanA participant or a real human wait point. People are never scheduled; a wait resumes with its original owner.
Agent / LeaderA logical role showing its real binding, current call and context. Model calls go through the same Brain Worker contract.
AI GroupAn expandable team with a named lead or an explicit summary rule. No invented Leaders.
WorkflowA process call, static or dynamic, referencing its definition and its actual Run on the shared engine.
IterationAn explicit loop node with a body, an exit condition and the history of every round.
Tool stepA real file, transform or script step on a specific Device or Hand.
Static, dynamic, runtime change and iteration.
Each is a distinct capability with its own acceptance. None of them impersonates another.
A plan the Leader proposes
Candidate flows generated from the goal and context, then started only after you accept them. A plan is not an execution.
A graph that grows with the facts
New facts can add, remove or replace steps that have not run yet. The source of the change, its diff, the accepted version and the affected scope stay visible.
A loop you can see
Repeat a body until a condition holds or a budget stops it. Hitting the round limit without meeting the exit condition is never shown as success.
Steer the run. Never rewrite the past.
Unexecuted nodes show as not started. A connection shows only when a real hand-off happened. A reservation is not Running. Every control goes through the same authorized, idempotent command path.
Confirmed, not assumed
Stop ends the current round through one idempotent path and waits for the device to confirm. Continue requests resources again instead of holding them idle.
Manual retry keeps history
A new attempt never overwrites the old result. Dependent downstream results are explicitly invalidated and re-run.
Ready steps run together
Parallelism follows dependency readiness, real capacity and actual write conflicts, not a fixed wave schedule.
Drill into any call
Every real invocation can be opened from the canvas: executor, inputs, outputs, artifacts, errors and usage.
Keep what worked
Processes and Skills that outlive the conversation.
A good run should not disappear with the chat. DAG Flow captures it as something you can open, inspect and run again.
Flows are files
A flow is a versioned definition with inputs, outputs and the AI or tools it needs. Export it, import it into another project, and run it on the same shared engine.
Skills in the Library
The Build workbench keeps Flows, Docs, Skills and Tools one click away. Roles such as researcher or data analyst are Skills, not new identities.
One engine, one truth
Authoring happens in Build (VLC). Orchestration, run views, cancel and human waits happen in Matrix. Execution belongs to the shared VL Base runtime. There is no second scheduler and no separate agent IDE.
Automatic template promotion and arbitrarily recursive sub-flows are outside the current release; existing flows and Agent Apps you have built keep running.