Enterprise Governance for AI Agents: The Control Layer Has Arrived

What happened in the last 72 hours

In just three days, four major vendors independently sent the same signal. On August 19, UiPath launched Maestro Flow, a developer-first orchestration canvas that gives coding agents such as Claude Code, Cursor and GitHub Copilot enterprise-grade durability and governance. The same day, Salesforce published data showing how fast the underlying problem has grown: customers' AI agent 'headcount' tripled between February 2025 and April 2026. On August 20, GitLab shipped version 19.3, introducing usage caps for GitLab Credits, group-level visibility restrictions for custom agents, and a production-ready Secrets Manager. And on August 21, Google folded Antigravity into Gemini Enterprise subscriptions, adding admin controls for licenses, security and spend, sandbox restrictions, and approval gates for terminal commands.

• Google Antigravity (Aug 21): admin controls for licenses, security and spend, sandbox restrictions, mandatory approval for terminal commands — now part of Gemini Enterprise.
• UiPath Maestro Flow (Aug 19): a developer orchestration canvas connecting coding agents to enterprise-grade durability and governance.
• GitLab 19.3 (Aug 20): usage caps for Credits, group-level visibility restrictions for custom agents, general availability for the Flow Creator Agent, Secrets Manager support.
• Salesforce data (Aug 19): agent 'headcount' tripled in 15 months, time to activation down 53%, average actions per account growing at a 31% CAGR per month.

Why this is more than a feature update

The pattern is unmistakable. After a year in which enterprises were largely occupied with agent sprawl — fleets of agents multiplying without central oversight — the major platforms are now delivering the technical answer. The question has shifted from 'how do we build an agent' to 'how do we keep a hundred of them under control.'

This control layer has been the missing piece between a working agent prototype and a production-grade system. The fact that four vendors moved within 72 hours of each other shows that the market now treats governance not as a nice-to-have, but as the precondition for letting agents anywhere near production.

What the new control layer actually delivers

• Centralized administration: licenses, cost and security policy for every agent in one place instead of scattered across dozens of tools.
• Usage limits: hard caps on resource consumption so a single agent cannot silently burn through budget or compute.
• Visibility and approval rules: who can see, change or run which agent, and under what permissions.
• Sandbox boundaries: agents operate in restricted environments by default, with critical actions requiring explicit human approval.

This lines up with an observation now surfacing in almost every enterprise rollout: the organizations creating real value with AI agents are not the ones running the most agents — they are the ones with the clearest limits on what an agent may decide alone.

Enterprises winning with AI agents are limiting how much the agents can do alone.

Why this matters even more for Swiss enterprises

For Switzerland-based organizations, the business logic is reinforced by regulation. When an agent makes automated decisions, Art. 21 of the revised Data Protection Act (revDSG) applies, with requirements for traceability and the right to contest a decision. When a bank or financial services firm deploys AI agents for outsourced processes, FINMA Guidance 08/2024 on outsourced AI applications comes into play. And under Art. 716a of the Swiss Code of Obligations, the board of directors carries a non-delegable duty of ultimate oversight — including for agents acting autonomously in the background.

From tool to lived governance

An admin dashboard with usage limits is a tool. Governance only emerges once an organization defines which decisions an agent may make autonomously, which require sign-off, and who is accountable when something goes wrong. That decision logic differs substantially between platforms, which is precisely why the choice of control layer is also a choice of operating philosophy.

The next step for your organization

The vendors are now delivering the building blocks. What they cannot deliver is the company-specific answer to what comes next: which agents you actually need, where the boundary of autonomy sits, and who runs this control layer day to day without it becoming yet another internal project competing for capacity. This is exactly where an external AI division earns its place — treating governance not as a one-off project, but as an ongoing operation.