Govern your AI agents.
Our specialty, and our only one: AI agent governance — writing the rules, testing, proving. From the organisation with no policy yet to the agent that has to reach production.
Agent governance.
Anyone can plug in an agent. The real challenge starts after: an agent that acts autonomously is a new kind of risk. It decides, it executes, it touches your data and your clients — and without oversight, no one can explain or defend what it did. That's what we govern.
We put in place the layers that make an agent controllable — built in from the design stage, never bolted on afterward. The result: an agent you can genuinely deploy, control and defend before your leadership, your auditors and your stakeholders.
Controls and guardrails
Clear limits on what the agent can do, and what it can't.
Security & agent risk
We identify and contain the risks specific to agent autonomy.
Human oversight
A human in the loop at decisive moments, with the right to review and to stop.
Reproducible proof
We stress-test the agent adversarially with an open harness (Apache-2.0): your auditor can re-run the exam without us.
Continuous monitoring
We track agent behavior in production to catch drift.
Auditable traceability
Every decision is logged and defensible, aligned with your sector's requirements.
Testing runs on the ProofAgent platform (ProofAI LLC), with which Aurelia Ops has a commercial relationship, and on its open-source harness (Apache-2.0). The method is written down and the exam can be re-run by your auditor — that is where independence lives, not in the status of the engine.
Your rules, written and tested in 90 days.
Your teams already use agents, build them or buy them. But the policy that says what they may do, who answers for them and how you prove it does not exist yet — or fits in three principles nobody can apply.
We write it with you, in 90 days and twelve steps: see where you stand, decide what you protect and what you allow, write rules that each have an owner, a control and a proof, then run them on a first agent.
The path follows the book 90 Days to AI Agent Governance Readiness, which we co-wrote with the founder of ProofAgent. You leave with a policy, a registry and a first proof file.
- Starting point. A workshop: your maturity level out of five, the inventory of your agents — used, built, bought — and your fears and ambitions put on the wall.
- Policy v1. Three working sessions: sort the scenarios, write the rules, fit on one page — nine blocks, from mandate to exceptions and incidents.
- Agent registry. Every agent, its risk tier, its gate and its next review.
- First agent under test. Your rules loaded into the platform, one agent put to the test, its proof file in hand — and the version 2 of the policy that follows.
Your agent, tested before production.
An agent stuck in pilot because nobody dares sign off on production. An agent a bank, a client or a regulator wants to see approved. An agent already in production whose auditor asks for evidence. In all three cases the question is the same: what does it really do, and can you defend it?
We test it adversarially, with an open harness your auditor can re-run without us, and we link every finding to the text that requires it — AMF, OSFI E-23, ISO/IEC 42001, OWASP, NIST, the EU AI Act, depending on your exposure.
You leave with a re-examinable proof file, and evidence kept current at every material change.
- Scoping. The agent's scope, its attack surfaces, the applicable frameworks and the test plan — written before the test.
- Adversarial testing. Injections, tool misuse, data leaks, mandate overreach: the agent is attacked the way it will be in production.
- Proof file. Findings, exhibits, mapping to requirements, limits of the exam — the document your risk committee and your auditor can read.
- Proof keeping. New model, new tool, new scope: the previous evidence no longer holds. We re-test and keep the file current.
To a governed agent in production.
Discovery call
One call to name the situation: rules to write, or an agent to test.
Starting point or scoping
For the organisation: a workshop that measures where you stand and inventories your agents. For the agent: its risk scope and the applicable frameworks.
Programme or file
90 days to write your policy and test a first agent — or the adversarial test and proof file of one agent, kept current at every material change.