Three questions, answered continuously
Without a control plane these answers live in scattered logs, static registers and individual tools. Under one, they are a single coherent picture that can be enforced.
What agents exist. A complete, current inventory of every agent, registered or not, built from the calls they make rather than from a register kept by hand. The control plane cannot govern what it cannot see.
What each may do. Policy tied to identity, purpose and real behaviour, so every agent has a defined boundary rather than whatever access it happened to be handed on day one.
What each actually did. A record of every call, decision and outcome, which is the difference between asserting that agents are governed and being able to show it.
A decision is not an enforcement
Many tools stop at recommendation. They flag a risky call, log it, and let it proceed. For a regulated enterprise that is commentary, not control. A control plane sits on the path of the call and allows or refuses it in real time, so the decision has teeth because it happens before the action does.
Any agent, any framework
On one side, the agents: personal assistants, LangChain and CrewAI, Claude and OpenAI agents, MCP servers and tools. On the other, the governed services: internal service agents, enterprise APIs, MCP tool servers and external SaaS, proxied.
Between them sits KPATH’s policy enforcement point, carrying identity, kill switch, cost, policy, approval and guardrails, and writing a tamper-evident audit to your SIEM.
One path between every agent and every system it can touch. Agents carry no system addresses and no credentials, so every call is identified, checked against policy and recorded, whatever framework the agent runs on.
Discover, identify, govern, contain, prove
KPATH structures agent governance as five capabilities. Discovery gives you the map, enforcement gives the map consequences, and proof turns the whole thing into something you can defend.
- Discover. Build a complete inventory by watching real calls, including agents nobody registered. Everything downstream depends on having the full map first.
- Identify. Attach identity and ownership to every agent, so each call has an accountable source rather than an anonymous credential.
- Govern. Set what each agent is allowed to reach and do, with policy grounded in observed behaviour rather than in assumptions about how an agent should act.
- Contain. Enforce those decisions on live traffic. Calls are allowed or refused at the point they are made, so risk is bounded rather than merely noted after the fact.
- Prove. Produce the evidence that shows what ran, what was allowed, what was refused and why, in a form a supervisor or auditor can accept.
One layer, four concerns that would otherwise be four programmes
Cost. Agent calls translate directly into spend. Seeing and governing them keeps costs from climbing without anyone deciding they should.
Governance. Policy applies consistently across every agent and framework, rather than being reinvented team by team with different limits each time.
Security. Over credentialled and unrecognised agents are contained at the point of the call, so the blast radius of any one of them stays bounded.
Observability. Every action is visible and recorded, so the questions that follow an incident or an audit have answers rather than gaps.
It works with the stack you already run
A control plane should not force a rebuild. Keep your identity provider, your guardrails engine, your agent platform and your existing APIs. KPATH observes both ends of every call across all of them and enforces policy in the path. It starts in monitor mode, so you see the full picture and what policy would do before anything is refused.
Related: AI agent discovery, or what the EU AI Act asks of agent estates.
Agents are already doing business. Govern it, and prove it.
Begin in monitor mode. Every call is permitted while you watch what the policy would decide.