Destaris
Browse the docs

Deterministic-first

Why most steps should be plain and predictable, with AI added on purpose.

Most automation should be cheap, reliable, and auditable. Destaris is built around that: the default step is a plain, deterministic one, and intelligence is something you add to a specific step on purpose.

What "deterministic" means here

A deterministic step does exactly the same thing every time, given the same input: an HTTP request, a data transform, a branch on a condition, a cache check. There's no model, no temperature, no surprise. Because nothing is inferred, these steps are:

  • Free to run — there's no AI cost, so deterministic runs read $0.
  • Auditable — you can read the workflow and know precisely what it will do.
  • Reliable — the same inputs always produce the same outputs.

The boring 90% of real automation — moving data between systems, filtering, formatting, conditionally acting — is exactly this kind of work.

AI is opt-in, per step

When a task genuinely needs judgment — classify this, summarize that, decide which of these actions to take — you add an AI step. It's a single node you drop in deliberately, not a default that colours everything.

This keeps the cost and the unpredictability of AI contained to the steps that truly need it, and leaves the rest of your workflow fast and free.

The cost of a step that has to look things up

The $0 above is only half the saving. The other half is what a deterministic step spares the AI steps around it.

An AI step is a conversation, and every turn re-sends the whole conversation. A step that spends twenty turns working out which files it should be looking at is billed for twenty turns' worth of context, not for the answer. Caching makes each turn cheaper; it does nothing about the number of turns.

So the cheapest AI step is the one that starts already knowing. A task.command that computes the file list, the diff, or the name of the failing test costs cents and removes the turns an AI step would otherwise spend discovering the same thing.

On a real feature-delivery Loop, a successful run averaged 60M tokens — and almost none of that was new content. It was AI steps re-reading, turn after turn, what an earlier deterministic step already knew.

Determinism is a first-class signal

Destaris surfaces determinism in the product, not just the philosophy:

  • Deterministic steps show $0; AI steps carry a Uses AI chip.
  • Node categories are colour-coded on the canvas, so you can see at a glance which parts of a workflow are deterministic and which lean on a model.