Workflow basics

The model

A workflow in ThorStack is a directed acyclic graph (DAG) of nodes. Each node is one of:

• An agent step — invoke a specific agent with a specific tool.
• A human step — an approval gate or a hand-off.
• A branch — conditional routing based on the run's state.
• A wait — a pause for an external event or a timer.
• A fan-out / fan-in — parallel execution that converges.

Workflows have a trigger: a schedule, an event from a module, a webhook, or manual.

Running a workflow

A run is one execution of a workflow. Every run carries:

• Start and end timestamps.
• The trigger payload.
• A trace of every node — inputs, outputs, errors, retries.
• The full memory deltas (what was read, what was written).

You can replay a run from any node, which is invaluable when iterating.

The editor

The visual editor lays nodes out as a graph. Connect outputs to inputs by dragging. Each node has an inspector panel for its inputs, retry policy, and gate (if any). Workflows are versioned — saving creates a new version, not an in-place mutation.

Validation

Before a workflow can be enabled, ThorStack validates:

1. The graph is acyclic.
2. Every required input on every node has a source.
3. Permissions on each tool call are within the workflow's scope.
4. Approval gates exist for every action class that requires one (see Approval gates).

A validation failure shows the offending node inline, not as a modal you have to dismiss.

Next

• Conditional steps — branching and expressions.
• Approval gates — human checkpoints.
• Event triggers — start workflows from module events.