Most teams govern AI agents after the incident, which is the one moment governance cannot help. By then the decision has been made, the data has been read and the action has been taken on the business's behalf.
This is the web edition of a short Mobilio Asia whitepaper on what changes when software stops answering and starts acting, and what to put in place before the number of agents grows. The designed seven page version is readable on our LinkedIn page.
The problem: agents act, but most governance still watches users
Access management answers a familiar question: who signed in. It was built for people. It does not answer what an autonomous system decided, which data it read, or which action it took for the business.
That gap shows up in ways most organisations recognise once they look for them. Decisions that cannot be traced back to a policy. Models running in production that nobody formally approved. Prompts that drift from version to version with no record of what changed. No audit trail to hand a regulator. And governance tooling that keeps monitoring users while agents act on the business's behalf.
Why it is hard: controls sit where the risk is not
- A policy document cannot stop a call that has already been made. Rules written down somewhere are not the same as rules applied on the request path.
- An agent holding a provider credential can route around any console. If the agent has the key, it can reach the model directly, and enforcement becomes optional.
- Detector severity is evidence, not a decision. A detector can say something looks risky. Deciding what happens next is a policy's job.
- Zero traffic reads as a pass unless something says otherwise. A control that saw nothing has not proven anything.
- Every framework and runtime speaks its own dialect of the same problem. Without a shared vocabulary, each team governs its agents differently.
What good looks like: every action decided before it happens
- One identity, one owner and one lifecycle for every agent. An agent nobody owns is an agent nobody will switch off.
- One runtime vocabulary. Allow, block, sanitize, require approval and log only, applied the same way everywhere.
- Findings from detectors, decisions from policy, never the reverse. Detection informs the decision; it does not make it.
- Evidence kept without keeping the prompt or payload. The record shows what was decided and why, without storing the content itself by default.
- Readiness reported and scoped, never certified. Readiness belongs to a stated use case, jurisdiction and deployment stage. It is not a claim that a system is compliant.
How Mobilio approaches it: log only first, enforce on purpose
This is the order we use on our own rollouts of Syncalytics, the AI governance platform Mobilio Asia builds and operates:
- Register the agents that already exist. Start from what is running, not from what is planned.
- Put a governed route in front of the model, so the provider credential never sits inside the agent and enforcement cannot be bypassed.
- Run guardrails in log only mode first, and read what they would have decided before they decide anything.
- Switch on enforcement one scope at a time, with a named owner for each.
- Choose the failure behaviour deliberately. Fail closed for production and regulated work.
What to ask a vendor: six questions, including for us
- Does the control sit on the request path, or does it read logs after the fact?
- Can an agent bypass enforcement by calling the model directly?
- What happens to a request when a detector fails?
- How is zero traffic reported: as a pass, or as no data?
- Is raw prompt or payload content stored, and who can read it?
- Does the product claim to make us compliant, or does it report readiness scoped to our use case?
Where to start: the agents you already run
Most organisations have more agents in production than their register shows. The first step is to find them, give each one an owner, and put them behind a governed route in log only mode. Enforcement can follow once the team has seen what it would decide.
If you would like to talk it through, write to contact@mobilio.asia. Mobilio Asia works from Singapore, Ho Chi Minh City and Hanoi.