Deterministic automations next to AI agents

When the steps never change, write them once.

A script is a short piece of JavaScript that calls the same tools your agents use, with no model in the loop. The same steps every time, no tokens spent, and the same approvals, run history and schedules as an agent.

Soon Built, and not released yet.

The problem

Not every job needs a model to decide.

Some agent jobs are the same steps every time: read a sheet, post the row for today, send the reply that matches. A model adds cost and a chance of a different answer, and decides nothing. Writing it as code elsewhere loses the approvals, the run history and the schedule your agents have.

How it works

Code that calls your tools, run like an agent.

A script is JavaScript with your enabled tools. It runs in a sandbox, with limits on time, memory and calls, and no access to the network or the file system except through those tools.

  • No tokens. No model is called, so a run costs nothing in tokens and returns the same result for the same input.
  • The same gates. Any tool can require a person's approval; the script waits for the decision and continues where it stopped.
  • The same record. Every run, tool call and log line is on the timeline, like an agent's.
  • Works with agents. An agent can call a script as a tool, and a script can hand work to an agent.
Agents and scripts in one pipeline, one timeline.

Agent or script

Use a model where judgement is needed, and code where it is not.

An agent

  • A model decides each step.
  • Handles the unexpected: new wording, missing data.
  • Costs tokens on every run.

A script

  • Your code decides each step.
  • Does the same thing for the same input, every time.
  • No model, no tokens.

Templates

Start from an agent that already works.

45 templates for support, email, operations and reporting: install one, connect its tools, and gate the step that matters.

Solutions

Give agents a job, not a demo.

Support, sales outreach, social posts and operations: what the agent does, where a person decides, and what to watch.

Give one job to an agent this week.

Start with one repetitive workflow, gate the sensitive step behind your approval, and read the first run end to end.