When Caddie sets something up for you, it isn't promising to remember. It's compiling your ask into a workflow — a precise, executable definition that runs on the B3OS engine, independent of any AI in the loop.

That distinction is the difference between an agent that seems reliable and one you can leave alone with real money.

What "Deterministic" Buys You

  • Validated before live. Every automation Caddie builds is checked against the platform's schemas — actions, triggers, connectors, wallets — before it can be published. Malformed automations don't ship.
  • Runs without the model. Once published, the engine executes the definition directly. The AI helped author it; it doesn't improvise at runtime. The same input produces the same behavior.
  • Versioned. Edits create new versions with the old ones retained, so changes are reviewable and reversible.
  • Observable. Every run is recorded — what triggered it, each step's inputs and outputs, what it cost. See Monitoring.
  • Retryable. Failed runs are inspectable and re-runnable. Caddie can read a failed run and explain what went wrong.

Same Engine, Two Doors

There's one engine with two ways in: chat with Caddie, or build directly in the visual editor. Both produce the same artifact. Builders who want full control should start at Workflows and Architecture.

Why It Matters for Trust

The confirmation gates in Money & trust are enforced at this layer — not by prompt instructions, but by the execution engine itself. An automation's abilities are fixed by its reviewed definition, which is why "it only does what you approved" is a guarantee rather than a hope.