Cognitive Fab › Workflows
Workflows · the plan as a machine

The plan, admitted
before it runs

An agent that rebuilds its own to-do list every run will rebuild it differently every run. A workflow is that plan as a machine — admitted only if every path through it passes, then executed durably, one work order at a time.

Workflows are the deterministic form of re-planning. The sequence an agent follows every Tuesday was always a machine; leaving it to inference just means paying for it again on Wednesday and getting a different one.

What is being replaced

today · re-planning

Every run starts by reasoning out the plan again. Two runs of the same job take two different routes, an interrupted run has no state to resume from, and a step that fired twice is discovered from its effects.

instead · an admitted machine

The plan is a state machine that had to pass a check before it was allowed to load. It runs durably: one commit per step, effects emitted exactly once, and a journal that says who caused each step. Several agents can share one run without doing anything twice.

Admission, not advice

A workflow is model-checked before it is admitted. If a path through it can reach a state your rules forbid, it does not load — and you get the sequence that would have got there.

The agent reasons about the workflow

Not about the next tool call. The engine hands back one work order at a time; the agent does the work. The engine holds no credentials and calls no tool itself.

Exactly once, and resumable

Durable execution means a run that dies mid-flight resumes where it stopped rather than starting over, and an effect that already fired does not fire again on the retry.

A crew, not a swarm

Several agents claim work orders nobody assigned them and never take the same one twice. Identity comes from the process, never from a tool argument, and the journal records who caused each step.

The marks

Both open source under Apache-2.0, and free.