Enterprise agents need an approval loop, not just logs.

The 2026 agent platforms have converged on one promise: visibility. Traces, spans, evaluation scores, cost dashboards. Visibility matters — but it is fundamentally reactive. You see what the agent did. After it did it.

An enterprise, though, does not run on visibility. It runs on segregation of duties. The person who requests a payment is not the person who approves it. That principle predates software by centuries, and no dashboard replaces it.

The pattern we kept seeing

When we mapped the field — ServiceNow's AI Control Tower, Microsoft's Agent 365 "control plane", LangChain's Govern stage — the honest summary is: governance is being productized as observation and policy surfaces. You can see everything, and stop almost nothing before it happens.

To be fair, that is a real advance over raw logs. But compare it with what the same vendors' own enterprise customers already require for humans: requests are staged, approvals are multi-node, duties are segregated, and every decision lands in an append-only record. Agents should be held to the standard companies already meet — not a lower one because the actor is software.

What an approval loop actually is

The mechanism has three frames, and every frame is enforceable in code:

• Propose. The agent never writes directly. It produces a staged proposal against a typed action — which objects, which fields, which rule set attached.
• Approve. A multi-node workflow with segregation of duties enforced by the engine: initiators cannot approve their own proposals. Volume is handled by batch decisions; span by delegation.
• Audit. Every decision — approve, reject, return, delegate — lands in an append-only, bi-temporal trail keyed by causal ID. You can replay any decision and ask what the world looked like when it was made.

Nothing in this list is exotic. It is how ERP systems have treated sensitive human actions for thirty years. The only new question is whether we hold agents to the same standard.

The objection, and the answer

"Approval loops add latency." Yes — on the write path, deliberately. The fix is not to remove the gate; it is to make gates proportionate: read-side actions flow freely, low-risk writes get batch approval, critical control points hard-stop. Governance is a policy dial, not an on/off switch.

Where Kernos stands

Kernos ships this loop as its core: propose, approve, audit — every time, on a platform you self-host so the trail stays inside your perimeter. We built it this way because "log it and move on" is exactly how enterprises end up banning agents altogether.