ControlAutonomy
Autonomy
you can prove.
A guardrail classifies every tool call before it touches the outside world: execute, check, ask a person, or refuse. No model sits in that decision, so it cannot be talked out of it.
What it does
What control do
01
Deterministic guardrails
A JSON Logic rule sorts each proposed call into execute, execute if a guard passes, ask a human, or refuse.
02
The strictest wins
Attach guardrails to agents and tools; when several apply, adding one can only tighten the outcome.
03
Approvals with the reasoning
A held call arrives with the agent’s reasoning, evidence and predicted impact, so a reviewer decides on facts.
04
Exceptions and audit
What a guard stopped is recorded as an exception, and every write, refusals included, leaves an audit entry.
Guardrails, approvals and exceptions from Pro; the audit log from Business. Compare plans
Tutorial
Gate a tool with guardrails
By the end you will have an agent whose refund tool is governed: small refunds run on their own, large ones wait for a person, and you approve one and watch the agent finish. Every step is one call, shown for the CLI, the SDK and curl.
Start the tutorialMore features
More features
- AgentsA versioned configuration with tools, knowledge and a model.
- KnowledgeGround answers in your own documents, with the source to cite.
- ChannelsPut an agent on WhatsApp or Discord, where people already are.
- AutomationRun agents on a schedule, an event or by hand, with a record of each firing.
- EvaluationsScore a change against your own cases before it ships.
- ModelsManaged models with no vendor account, or your own credential.
Build the agent.
Keep the receipts.
Sign in, connect your assistant or open the quickstart, and run your first agent.
Free plan · 2,000 runs a month · no card
