Every action checked before it runs. Every decision on record.
AI agents are easy to demonstrate. Making them reliable is harder. We design and review the verification and governance layer that sits between an agent's reasoning and anything that touches money, data or customers. That covers new systems we build and systems you already run.
Secular Labs builds governance into the architecture, not a policy document. Every action an agent proposes is checked by deterministic rules before it runs. Every state change is written to a tamper-evident log. Repeated failures go to a person, and agents never hold payment or system keys. The approach is formalized in the AMEG (governance) and AMES (security) papers.
The controls
| Control | What it prevents |
|---|---|
| Deterministic orchestration | Agents drifting off-task in open-ended loops, and unpredictable cost |
| Verification-gated execution | Wrong, malformed or non-compliant actions reaching money, data or customers |
| Bounded self-correction | Endless retries, or a bad answer “fixed” into a worse one |
| Human exception handling | Agents guessing their way through edge cases |
| Auditable execution | “Why did it do that?” with no answer |
| Zero-trust boundaries | Prompt injection and stolen credentials: agents never hold keys, and independent security checks run on every action |
How it supports EU AI Act obligations
| Article | Typical practice today | With this architecture |
|---|---|---|
| Art. 9 Risk management | Periodic risk spreadsheets | Rules checked before every action, continuously |
| Art. 11 Technical documentation | Manuals updated now and then | Versioned architecture tied to every build |
| Art. 12 Record-keeping | Unstructured chat logs | Tamper-evident record of every state change, tool call and check |
| Art. 14 Human oversight | Reviews after something goes wrong | Automatic hand-off to a person when checks keep failing |
| Art. 15 Accuracy, robustness, cybersecurity | Periodic scans | Agents hold no payment or system keys; independent security checks on every action |
This shows how the architecture supports these obligations. Whether a specific system is high-risk, and how it complies, is assessed for each engagement.
What we can do for you
- Review an existing agent system: where checks are missing, where an agent holds more access than it needs, and what's logged and what isn't.
- Design the verification layer for a system you're about to build.
- Build it, using the same open-source policy, logging and observability tools we use ourselves.
Limits we state openly
| Limit | How we handle it |
|---|---|
| Checks catch rule-breaking actions, not actions that follow the rules but miss the intent | A named owner reviews exceptions and decides what agents may not decide |
| The safety bound assumes checks fail independently | Checks use different mechanisms, and repeated failures always go to a person |
| Monitoring after go-live | Error, escalation and cost rates are tracked, with re-validation when a model, prompt or vendor changes |
The research behind it
AMEG covers governance, accountability and audit. AMES covers zero-trust security between reasoning and settlement.
Frequently asked
How do you verify an AI agent's decision?
By not letting the decision commit until it passes deterministic checks such as schema, business rules and signatures. Each check is independent of the model. Failures are stopped and sent to a person.
Does this make a system EU AI Act compliant?
It provides the controls the Act asks for, such as record-keeping and human oversight. Whether a system is high-risk, and how it complies, is assessed per engagement.
Can you review a system someone else built?
Yes. A review covers access, checks, logging, hand-off and cost visibility.
Do agents ever hold credentials?
No. Payment and system keys stay outside the agent runtime, and only the final settlement step can use them after every check has passed.
Where do we start?
With the free diagnostic or a 30-minute call.
Start with one workflow.
Take the free diagnostic (8–12 minutes), or book a 30-minute call. The Operational Leverage Audit starts from $1,500; a Pilot on one workflow starts from $6,000. Each step can be the last one.
Related services: AI-native operating model · AI workflow automation · AI systems due diligence · Applied systems · Explainer: Verifying AI agent decisions