Concept
Turn AI on for one step. Not your whole stack.
AI everywhere is a liability; AI where you choose is a tool. In Destaris, an AI step is something you opt into deliberately — pick the model, keep it contained — never the engine the whole workflow quietly runs on.
One node, opted in
AI is a setting on a step — not a mode for everything.
Turning on AI for a step is a choice you make on that node, the same way you'd pick a URL for an HTTP step or a schedule for a trigger — not a switch that changes how the whole workflow runs.
Ask AI
agent.run
Model
Effort
Contained by design
One step's blast radius stays one step.
An AI step takes typed input and returns typed output — nothing more leaks in or out. Everything upstream and downstream of it keeps running exactly the way it always has: deterministic, inspectable, and free.
Everything outside this boundary stays deterministic — plain steps, same every run.
The step reads a typed in, hands back a typed out, and the rest of the graph decides what happens next — the model never gets a say beyond the one node you gave it.
Pick the muscle
Match the model to the moment.
Fast
For quick, cheap checks and classification — routing an enquiry, tagging a message, a yes/no gate. Cheap enough to run on every step that just needs a fast read.
Balanced
The everyday default — drafting a reply, summarizing a thread, most of the judgment calls a workflow needs. Reach for this first; upgrade only when it isn't enough.
Most-capable
For the hard reasoning — multi-step decisions, ambiguous input, anything where getting it wrong is expensive. Slower and pricier, spent only where it earns its keep.
Go deeper on AI opt-in per step
Read the concept doc for the full model — when to reach for an AI step, how it runs on your own already-authenticated CLI, and how to keep its output typed and contained.