ACT and Ground Control: real-time governance for enterprise AI agents

Prepare the company for a future with many AI agents at work in a regulated, US-listed oil and gas environment, where a data leak or an unchecked failed solution carries legal and reputational cost.

Goals:

1. Know every agent. A central registry holds each agent's objective, owner, guardrails and token limits before it starts working. This replaces inventories compiled by hand.

2. Control access. Deterministic code authorizes every request to APIs, MCP servers and systems against company permission policies.

3. Stop any agent. Every agent exposes a halt callback, so an operator can stop a runaway or compromised agent from one control room.

4. Monitor consumption. Tokens are tracked by day, week, month, quarter and year, with alerts near limits.

5. Stay independent of any model vendor, so governance works with any platform.

Status, October 2026: ACT and Ground Control are built and operating.

2026

October 2026. ACT and Ground Control (since September) are operating. Five people delivered them between June and October: a tech lead, an architect, two developers and the head of AI.

Readings: 4 agents are registered in ACT. ACT audits about 37 events per day. Ground Control recorded 23 access denials in its first two weeks. The live registry replaces manual inventories.

Roadmap: durable access logging in Ground Control, registration of every agent by December 2026, migration of GPTs to skills and plugins, and testing the protocol for user GPTs and skills. It is written as a company-wide skill, and Halt sets a user's usage availability to 0 or deactivates the user through the enterprise management console API.

Python (ACT and Ground Control services), React (operator control room), REST APIs, MCP servers, agent-to-agent (A2A) architectures, model provider APIs (OpenAI and Anthropic), custom GPTs, skills and plugins on a company-wide Claude deployment.

The project does not use solutions from the Global CIO catalog

Governing agents of different technologies with one protocol: custom GPTs, skills, plugins and code agents each needed their own path to register, report and expose a halt callback.

Translating existing permission policies into deterministic rules that Ground Control applies to every API, MCP server and system.

Adoption: agents built by users and by developers must register and expose a halt callback, which depends on every builder.

Changing assistant platforms and pricing models: the governance layer works with any of them.

Effort: five people delivered both services in five months.

Auditability: ACT keeps audited events. Ground Control access logging is the next piece to build.

Initial situation. Since May 2024 the company has built AI assistants for employees: 80 custom GPTs backed by internal services, 12 of them with high use. More than 1,500 further GPTs have no connection to company services, spend tokens from the company subscription and have no live inventory. Since June 2025 a team of 10 automation hackers (developers who build with AI) has also built agents in code, using agent-to-agent architectures, MCPs and APIs. About 400 employees use these tools. The number of agents is set to grow, and each one can reach data, spend tokens and be manipulated. For a regulated company that reports to a US stock exchange, one information leak or one unchecked failed solution can carry heavy consequences. Agent inventories were compiled by hand, and no mechanism existed to stop an agent.

Implemented solution. Two in-house services apply the governance policy as code.

ACT (Agent Control Tower) is a REST API with four functions. Register: an agent declares its name, objective, guardrails, creator, token limits and halt callback, then receives a key. Report: lifecycle events (action started, in progress, completed) flow to the tower. Alert: guardrail bypass attempts and approaching token limits raise alerts, with consumption tracked by day, week, month, quarter and year. Halt: the tower calls the agent's halt callback to stop a rogue or compromised agent.

Ground Control is deterministic code. An agent presents its key and asks for the list of MCP servers and APIs it can use. Ground Control confirms with ACT that the key is registered, checks it against company permission policies and issues a temporary token for the allowed service. An unregistered agent receives no list and no access.

An operator control room built in React shows every registered agent on a radar, the access requests Ground Control allowed or denied, token consumption and halt status.

Implementation specifics. The services run in Python and agents join through MCP servers and API calls, so any agent technology can integrate. Five people delivered both services in five months.

Changes after implementation. Operators have one control room. A live registry replaces inventories compiled by hand, every registered agent has an owner, guardrails and token limits, and every denied access is recorded.

Global. ACT and Ground Control govern agents across all company operations.