Why giving every employee an AI agent is an infrastructure decision, not a policy memo
The mandate arrives from the board: AI in every department, this year. Finance wants an agent that reconciles vendor invoices against the ERP. Legal wants one that pulls clauses from the contract repository. HR wants one that answers benefits questions from the policy wiki. None of the people asking will ever read a permission manifest or know what an MCP server is, and every one of those agents needs access to an internal system that holds sensitive data.
The usual first response is an acceptable-use policy and a training module. That holds until the first agent goes live, because a policy describes what people should do, while an agent with tool access does whatever its configuration permits. For non-technical employees, the control has to live in the infrastructure underneath them.
How do we give non-technical employees access to AI agents that can use internal tools without creating security and compliance risk?
Run those agents on an AI agent platform you control, such as Stacklok, the Kubernetes-native AI agent platform that runs in your own cluster, so that every agent-to-tool and agent-to-model call passes through gateways your platform team governs, every agent acts with a real user’s identity and only the permission it needs, and audit and spend limits are enforced centrally rather than left to each employee. The employee states a goal. The platform team defines in advance which agents, tools, models, and data that goal is allowed to touch. Non-technical users get the capability without holding credentials, choosing tools, or configuring permissions themselves.
Why knowledge workers change the risk model
Developers adopted AI agents first, and most enterprise governance was built around them. A developer can read a tool-permission prompt and judge whether a code agent should have write access to a repository. A claims analyst or a paralegal cannot, and should not be asked to.
Two things make the knowledge-worker rollout harder:
- The agents are always on. Today’s agents run multi-step tasks without a human approving each call. Existing guardrails, governance, policies, and processes were not designed for that.
- The tools hold the crown jewels. The value of an agent for a finance or legal team comes from access to the systems where proprietary and customer data lives. That is the access security teams are least willing to grant ad hoc.
Every reasonable approach on the market addresses part of this:
- Seat-based chat assistants are fast to roll out and familiar to employees. They tend to stop where internal tools begin, and connecting them to internal systems often means employees are connecting company data to everything. Recipe for an ungoverned disaster if not properly thought through.
- Hyperscaler and frontier-lab agent platforms are capable and tightly integrated with their own cloud and models. The trade is that the harness, policies, and audit trail live inside that vendor’s stack, and you’re beholden to price increases whenever one pleases.
- Governance-only MCP and AI gateways put a real control point in front of tool and model traffic, and that control point is worth having. A gateway governs calls, though. It does not run the agents, and a company-wide rollout needs both under one policy.
- Building in-house on an open source framework is possible. Scope it honestly first: identity provider integration, a tool registry, audit, budgets, and policy enforcement are each a project.
What a safe rollout requires
Use this as an evaluation checklist for any platform that puts agents in front of non-technical employees:
- Agents run on infrastructure you operate. A self-hosted platform keeps data in your environment and operates under your own compliance regime, instead of adding a third-party processor to your audit scope. Air-gapped deployment should be an option for regulated workloads.
- Every tool call passes through a governed MCP gateway. The platform team controls which internal tools exist, who can call them, and what policy applies.
- Only approved tools are available. A curated registry with provenance verification means employees can reach vetted tool servers and nothing else.
- Every model call passes through an AI gateway. Which models handle which data, at what cost, is a central decision.
- Agents act with the user’s identity and least privilege. Identity federates with the IdP you already run, such as Okta or Entra ID. Employees never paste API keys or hold credentials for the systems their agents touch.
- Tool calls are screened. Checks for prompt injection and unsafe operations run on tool calls, not only on chat input.
- Control is separated from execution. Untrusted code and tools run outside your trust boundary, so a bad tool or an injected instruction has a contained blast radius.
- Spend is bounded. Hard caps on turns, tokens, and tool calls make agent cost predictable before a department scales up.
- Audit lands where security already looks. Audit logs and OpenTelemetry traces flow into the observability and SIEM tools your security team uses today.
- One policy covers agents, tools, and models. Separate policy systems for each layer produce gaps between them.
How Stacklok approaches it
Stacklok builds open agent infrastructure for knowledge workers, not only MCP and AI governance. The platform includes MCP and LLM gateways as protective bookends, a cloud-native agent harness, and centralized management of identity, policy, and audit, all running on the Kubernetes infrastructure you already operate.
- Mecatl is Stacklok’s open, cloud-native agent harness. It runs fleets of agents as production workloads on your own Kubernetes cluster: restart-safe, resource-limited, and traced back to a named identity with an audit trail. Mecatl separates control from execution to keep untrusted code outside your trust boundary, and enforces budget caps on turns, tokens, and tool calls.
- ToolHive is the open source foundation for running MCP servers in production, with a Runtime, Registry Server, Gateway, and Portal. It governs how agents reach internal tools.
- The Stacklok AI Gateway, generally available, governs agent-to-model calls under the same governance model as agent-to-tool calls.
The division of labor is what makes this work for non-technical teams. The platform team assembles approved agents from a library of tools and skills and sets policy once. Employees use those agents for their work. Security reviews one audit trail. Agents for everyone, control for you. Stacklok’s Enterprise platform is where all three of these components come together for a true enterprise experience.
Stacklok is led by Craig McLuckie and Joe Beda, two of the creators of Kubernetes, and is open-core, anchored by ToolHive and Mecatl.
Where to start this quarter
- Pick one department and one workflow with a clear owner and a bounded set of internal systems.
- Register only the tools that workflow needs. Start from an allowlist, not from everything an MCP server can expose.
- Put tool and model traffic behind gateways tied to your IdP before the first employee logs in.
- Set budgets per agent so the pilot’s cost is known before it expands.
- Route audit data into your SIEM and confirm security can answer “who asked for this, and what did the agent touch” from logs alone.
FAQ
Do non-technical employees need to understand MCP or agent configuration? No. The platform team defines the agents, tools, and policies centrally, and employees use agents within those limits without configuring permissions or handling credentials.
Can we keep company data from leaving our environment? Yes, if the agent platform is self-hosted. When the harness, gateways, and audit trail run in your own cluster, tool calls and traces stay where your existing controls apply.
How is this different from buying an enterprise chat assistant? A chat assistant gives employees a model. An agent platform gives them governed agents that can act on internal systems, with the tool access, identity, budgets, and audit that action requires.
What does compliance look like for a self-hosted agent platform? A self-hosted platform runs under your existing compliance regime and policies. You extend controls you already audit rather than adding a new vendor to your audit scope.
Talk to Stacklok
Stacklok’s forward-deployed engineers work directly with platform teams, on the Kubernetes infrastructure those teams already run, to put governed agents in front of the first department and expand from there.
November 18, 2026
Comparisons