Deterministic guardrails for runaway AI agents

A deterministic permission and action firewall between agents and your business tools

Problem

The more tools an agent can reach, the more likely it goes rogue: changing customer records, issuing refunds, triggering workflows, or leaking data after a prompt injection. Non-deterministic guardrails like 'LLM as a judge' collapse under injection. SaaS founders openly admit they have no clear answer for how traditional roles and permissions should govern dynamic agent actions.

Solution

A lightweight local or edge proxy that sits between agents and actual tool calls. It uses declarative rules instead of an LLM to define which tools each agent may call, allowed parameter ranges, rate limits, and approval thresholds. High-impact or suspicious calls are blocked by default and routed to human approval, with a fully reproducible audit log.

Users

SaaS teams wiring agents into business systems, internal automation owners, and developers building enterprise-grade agents

MVP

An embeddable SDK or reverse proxy compatible with major tool-calling protocols; declarative rule files for tool allowlists, parameter constraints, rate and amount limits; blocking high-risk calls into an approval queue; structured audit logs with a web viewer; a one-line CLI to wire into an existing agent

Revenue

Usage-based or per-agent subscription; open-source core rule engine with a paid team approval workflow and audit dashboard; enterprise tier adds SSO and compliance exports

Why now

Agents are moving from demos to real business actions, and authorization remains a clear gap that big vendors are unlikely to cover for small teams soon. Rule engines are mature, so a solo dev can ship a demonstrable guardrail layer in 4–6 weeks.