Guardrails,
by default.
Agents are powerful employees with no fear of consequences. Slaunt wraps every run in budget caps, scoped tools, human approval, and a kill switch that actually kills.
Five guardrails on every run
Not a policy document — enforcement. Each layer is on from the first run, and each one holds on its own.
At the cap the run pauses and asks. It never spends past it.
At the cap the run pauses and asks. It never spends past it.
Rules we don't bend
The guardrail stack follows four principles. They're design decisions, not settings.
Deny by default
An agent starts with nothing — no tools, no spend, no reach. Everything it can do is something someone granted.
Human in the loop
Guardrails don't slow work down; they decide which work needs a person. Everything else flows.
Kill means kill
A stop control that negotiates isn't a stop control. Kill halts the run — cleanup questions come after.
Written down
Every action lands in the audit trail as it happens. If it isn't logged, it didn't run.
Guardrails agents can read
Limits that live in a PDF get ignored. In Slaunt, an agent's scope is part of its context — it knows its caps, its tools, and its org boundary before it acts.
Every request an agent makes is bound to an authenticated identity and an organization scope. The agent can query its own limits the same way it queries workspace memory — which is why a Slaunt agent stops at a constraint instead of discovering it in an incident review.
Run agents you can trust
Budget caps, scoped tools, approval gates, and a kill switch — on every run, from the first one.