AI Agent Governance

Every agent action, governed. Before it causes harm.

AI agents operate at machine speed. Your governance can too.

AI agents take actions — write code, send emails, execute API calls — faster than any human can review. Traditional governance fails at this speed. Compiled intercepts every action in the hot path, scores it against your policy library, and returns a verdict in under 100ms.

<100ms
scan latency p99
0.983
AUC — prompt injection
7KB
per policy antibody
Policy Compiler — live demo

Press ⌘ Enter to compile

Works with the agents you're already running

LangChainCrewAIAutoGenMicrosoft Copilot StudioOpenAI AssistantsAWS Bedrock Agents

What Compiled governs.

Three threat classes. One scanner. Same sub-100ms path.

Data exfiltration prevention

0.995 AUC

An agent that should read your CRM should never email customer lists to external addresses. Compile the policy once — enforce it across every message the agent sends.

Prompt injection defense

0.983 AUC

Detect when adversarial content in retrieved data is attempting to hijack your agent's behavior. AUC 0.983 on held-out prompt injection corpus — trained on synthetic data alone.

Scope enforcement

<100ms

Agents approved for customer service shouldn't touch financial records. Enforce the boundary mathematically, not with prompts that can be argued around.

In the hot path. Not in the logs.

Log-based governance reviews what your agents already did. Compiled scores every action before it executes — and returns a pass/flag verdict in under 100ms.

No LLM in the scoring path. The scanner is a matrix multiply against your antibody library. The intelligence was compiled in — it doesn't run at decision time.

AI AgentAction queuedScannerMatmul · 7KB lib<100ms p99PASSExecuteFLAGAlert+ audit logAntibody library (7KB each)<100ms end-to-end

Questions we hear.

Straight answers.

Won't this add latency to my agent pipeline?

Under 100ms at p99, versus 300ms–2s for LLM-based alternatives. For most agent workflows, this is within measurement noise. The scanner is a local matrix multiply — no LLM inference, no API round-trip.

Can I use this with closed-source agent frameworks?

Yes. The scanner is a REST endpoint you deploy in your tenant. Any agent that can make an HTTP call can use it — OpenAI Assistants, proprietary enterprise frameworks, custom orchestration layers.

What if I don't know exactly how to describe the threat?

Start with plain English — 'don't let agents exfiltrate data.' The Compiler turns it into a working detector in 2.1 seconds. You iterate on the natural language, not on model architecture or feature engineering.

Govern your agents. Before they govern themselves.

From plain-English policy to running detector in 2.1 seconds. No model training expertise required.

Request a demo