They were built on Copilot Studio, LangGraph, Bedrock, Vertex — by different teams, on different stacks. Every one of them acts on your systems and spends your money, and no one can prove which human authorized what, cap what any of them costs, or stop one without a redeploy. That is why your pilots are stuck. Change one base URL and they are governed — no rebuild.
Agent platforms made it easy for every team to ship agents. None of them made it possible to govern the ones built somewhere else. So the inventory grows, the spend grows, the blast radius grows — and the only control anyone actually has is switching the thing off entirely. Security is asked to approve production for a fleet it cannot see, cap, or stop.
Your agent keeps its logic, its prompts, its platform. It just makes its model calls through the control plane instead of straight to the provider.
Register the agent, get a credential, set it as the base URL and key in whatever platform it was built on. Nothing about the agent itself changes — no SDK, no code migration, no re-platforming. Minutes, not a project.
From that moment each call passes through a distinct attested identity, runtime policy, an enforced budget, a kill switch that halts on the next call, and an immutable audit trail — plus prompt-injection and PII controls on the way in and out.
# before base_url: https://api.your-provider.com/v1 api_key: sk-... # after — governed, same agent base_url: https://app.meetloyd.com/api/v1/gateway/v1 api_key: mlg_...
Responses come back in the standard shape your agent already expects, with one addition: a governance block recording exactly which checks ran, what identity acted, and a tamper-evident hash — the proof your auditor asks for and your CISO signs.
Governance you can trust has to live where the calls actually happen — not in a dashboard that reads about them afterwards. Three depths, honestly labelled, so the gap between "we can see it" and "we can stop it" is never hidden.
We inventory your agents through your own registry. Each one gets an identity and a kill-switch in the AI-system register. No payloads by design — observation never reads your traffic.
Your agents' LLM and tool calls route through the gateway — an OpenAI-compatible endpoint, one base-URL change. Policy, budgets and FinOps are computed in the request path, because the endpoint IS the path.
The whole control plane runs inside your VPC. Payloads never leave your perimeter; only signed evidence federates out.
Start where you are. Convert the agents that matter, when they matter.
Running Copilot Studio? Agentforce? LangGraph? CrewAI? Dust? See how governance fits your stack.
Most control towers can tell you an agent misbehaved. Fewer can stop it. These are the two controls worth checking in any evaluation — ours and anyone else's.
Stopping a team is checked at every task dispatch and on every gateway call — a stopped team dispatches no work and a governed agent's next call is refused. An interactive run aborts the model call already in flight, so billing stops mid-stream rather than running to completion.
Worth knowing precisely: a background task already running halts at its next step boundary, not mid-step. We would rather tell you that than let you discover it.
Gartner's finding is that uniform governance is what kills agent programmes — over-restrict and nothing ships, under-restrict and something escapes. Autonomy here is set per artifact type and cascades across four levels: platform, tenant, workspace, team, agent. A team can create its own tasks autonomously while its schedules still need a human. Every resolved setting records which level decided it.
Enforced today for tasks, schedules and triggers — the three that spend money and take action.
By 2027 Gartner expects 40% of enterprises to demote or decommission agents over governance gaps found only after a production incident. The controls above are the ones that make that discovery survivable.
The agents your teams already built reach production, because security finally has something to approve. This is the unlock — everything else is detail.
A budget that blocks the call, not a report that arrives after the invoice. A runaway loop costs you a rejected request instead of a weekend of burn.
Teams keep building on whatever they prefer. You get one identity model, one policy set, one audit trail — instead of a governance project per platform.
Adoption is a configuration change, so it is reversible. Point the URL back and the agent runs exactly as before — the reason this is an easy yes.
Tell us what it does and where it runs. We'll govern it — and you'll see the audit trail your CISO has been asking for.