Practical AI agent governance: what running agents in production actually taught me

There’s a lot of conversation about AI agent governance right now. Frameworks, risk matrices, policy documents. Most of it is theoretical.

I’ve been running an AI agent in production for my software company SquareWave Studio for the past couple of months. It handles support triage, community monitoring, revenue tracking, content drafting. Runs automatically several times a day. Here’s what surprised me about governance when you’re actually doing it rather than planning for it.

Human approval on customer-facing actions costs less than you’d expect

Reviewing a drafted response takes seconds. Without that gate, a single bad response erodes customer trust that took months to build. The maths is obvious once you’ve lived it.

Audit trails earn their keep on the worst days

Something will go wrong eventually. When it does you need to know what the agent decided, what data it used and what it ignored. Building this in from the start took an hour. Retrofitting it after an incident is always harder and more expensive than building it in. Without observability, you’re debugging in the dark while stakeholders are asking questions you can’t answer.

Pre-execution gates on anything that matters

Money, orders, customer data, public content. Without gates on these you’re hoping nothing breaks. Hope is not a governance strategy.

The pattern is familiar

None of this is complicated. It’s the same lesson every technology wave teaches: build guardrails before you need them, make the safe path the easy path.

The EU AI Act enforcement starts in August. NIST published agent identity standards last month. SOC 2 auditors are already asking about autonomous workflows. The organisations building governance into their agents now won’t be the ones scrambling later.

The gap between “we’re experimenting with agents” and “we have agents in production” is mostly a governance gap. The technology is ready. The question is whether your controls are.

These lessons from running agents in production are part of what led me to build Vectimus. Every rule in the platform traces back to a real incident or a real operational lesson. Two commands to install, open source and compliance-mapped from day one.