Agentic AI acts.
MCP links agents.
Torque governs and reports actions.

Standard AI responds to prompts. It generates text, answers questions, and produces outputs for a human to review and act on. Agentic AI takes actions. It does not wait for human review between steps. It assesses a situation, determines what to do, and executes, potentially triggering further actions based on the result. That distinction is operationally significant. When an AI agent can provision infrastructure, scale compute, or terminate resources without human intervention at each step, the question of what governs those actions: who authorized them, what limits apply, who is accountable for the cost, is not theoretical. It is live, in production, right now, in most enterprise environments that have deployed AI tooling.

MCP (Model Context Protocol) is an open standard, developed by Anthropic and adopted across the industry, that enables AI agents to discover and invoke external tools and capabilities in a structured, governed way. When Torque implements an MCP server, any MCP-compatible agent can discover Torque’s capabilities and call them directly, without custom integration work. The agent learns what Torque can do: provision environments, query state, inspect drift, check costs, and can invoke those capabilities as part of its reasoning process. Crucially, the Torque MCP server enforces the same policy controls, cost limits, and audit trail as any other request. The agent gets governed access to infrastructure capabilities. The organization gets a complete record of what the agent did and why.

User roles govern what a human can do in Torque based on their job function and team membership. Agent roles govern what an autonomous agent can do based on its defined scope and purpose. They are separate objects in the platform for a deliberate reason. If an agent inherits user-level permissions through a service account, its blast radius is the same as that user, potentially very broad. Agent roles are scoped specifically to what that agent needs to accomplish its purpose and nothing more. An agent responsible for cost operations has read access to cost data and the ability to surface recommendations. It cannot provision or destroy. An agent responsible for deployment can provision within its quota. It cannot modify cost policies. The scope is enforced by the platform, not by the agent developer, which is the only way to make it reliable.

Without a coordination layer, this produces race conditions and conflicting state. A cost agent terminates instances the deployment agent just provisioned. A remediation agent and a maintenance agent make incompatible changes to the same environment simultaneously. Each agent is behaving correctly from its own perspective. Together they produce an outcome no one intended and no one can easily explain. Torque maintains authoritative state across every environment in the estate. When agent actions would conflict, Torque detects the conflict before it reaches infrastructure, applies the organization’s configured priority hierarchy: typically compliance first, then SLA commitments, then cost optimization: and resolves the conflict by sequencing, blocking, or escalating. Agent developers do not need to solve this individually. It is handled at the platform level, which is the only place it can work reliably at scale.

The Copilot works with whatever is in the inventory, but the quality and accuracy of what it generates is directly tied to the completeness of what Curate has discovered. If the inventory includes 200 validated IaC modules, the Copilot can generate precise, accurate blueprints from those assets. If the inventory is partial, the Copilot’s output will reflect those gaps, it may suggest generic templates for the missing pieces, or flag that certain components are not available in the governed catalog. This is not a limitation, it is how the system maintains the guarantee that AI-generated infrastructure is built from assets your organization actually owns and governs. Starting with a complete Curate discovery is the highest-leverage step an organization can take before activating the AI Copilot.

Any agent that can call a REST API or an MCP endpoint can use Torque’s governed infrastructure surface. This includes LangChain, AutoGen, CrewAI, and any other framework that supports tool calling or MCP protocol. It includes coding assistants like GitHub Copilot and Cursor when they are configured to use external tools. It includes CI/CD pipeline agents in GitHub Actions, GitLab CI, and Jenkins. It includes any custom agent built on any model, Claude, GPT-4, Gemini, or others. The governance layer does not care about the agent’s architecture. It applies policy, records attribution, and enforces cost limits to every request, regardless of the source.