The questions people ask about governing AI agents.
Plain answers to how Agentomy works, from the first governed action to compliance proof. If you are new, start with the first question; if you know the space, jump to what you need.
- What is AI agent governance?
- What is Agentomy?
- How do I set up governance for my AI agents?
- How do I govern MCP agents and tool calls?
- How do I monitor AI agents for drift or compromise?
- Can I stop an AI agent instantly?
- How do I prove AI agent compliance for the EU AI Act, SOC 2, or HIPAA?
- How do I test my agent platform against the OWASP Agentic Top 10?
- Does my data leave my environment?
- Which agent frameworks does Agentomy support?
What is AI agent governance?
AI agent governance is the discipline of answering three questions at the moment an agent acts, not after: was this action authorized, can the record of it be proven to a third party, and could a human have stopped it. In practice it covers authorization enforcement, a tamper-evident audit trail, behavioral monitoring, and an emergency stop, all applied before the action rather than discovered later in the logs. Agentomy is the governance layer that provides this for agents on any framework, vendor, or cloud.
- Where it lives
- Core concepts · The governance API, operation by operation
- How it is proven
- GovernanceBench, the open benchmark · Pattern provenance
- Where it stops
- Governance decides and proves what an agent may do. It does not improve the agent's model or replace the systems the agent acts on; it owns authority, not execution.
What is Agentomy?
Agentomy is a vendor-neutral governance layer for AI agents in regulated enterprises. It provides enterprise-wide discovery, policy enforcement, and compliance proof across every AI agent, model, and application, regardless of the framework or deployment. It runs self-hosted, so it governs the agents you already have without asking you to change them.
- Where it lives
- Platform · Machine descriptor
- How it is proven
- Published result manifest · Adapter capability profiles
- Where it stops
- Agentomy owns authority, not execution: it governs what an agent may do and proves what it did. The agent's own runtime still performs the action.
How do I set up governance for my AI agents?
Start with agentomy-agent, a free-forever governance peer you install in about thirty seconds with no account and no API key. It gives you agent discovery, permission limits, behavioral monitoring, and an honest posture score against GovernanceBench on your own machine. When you need cryptographic audit and enforcement you can show an auditor, you connect the self-hosted platform.
- Where it lives
- Deploy the agent · Install · CLI reference
- How it is proven
- What each adapter governs
- Where it stops
- Standalone mode assesses three of the six GovernanceBench dimensions on your own machine. Connected mode, with the self-hosted platform, assesses all six.
How do I govern MCP agents and tool calls?
The MCP Gateway interposes on any stdio MCP server and governs every tool call, failing closed on anything unauthorized. Because it sits in front of the server rather than inside your agent, MCP governance does not require changing your agent's code.
- Where it lives
- MCP Gateway · Gateway capability profile
- How it is proven
- Authorize operation in the API reference
- Where it stops
- The gateway decides every tools/call. Discovery (tools/list, resources, prompts) and server responses pass through unchanged, by design: what an agent may do is decided, what it may see is not narrowed.
How do I monitor AI agents for drift or compromise?
Agentomy runs behavioral monitoring on every action against each agent's own rolling baseline, watching for privilege probing, scope escalation, frequency spikes, and identity drift. When an agent starts behaving unlike itself, the action is flagged or the agent is halted. This is agent observability aimed at governance, not uptime.
- How it is proven
- GovernanceBench behavioural suite
- Where it stops
- Drift is judged against the agent's own rolling record, never another agent's. A brand-new agent has no baseline yet, so its first actions are governed by authorization and policy until the record exists.
Can I stop an AI agent instantly?
Yes. A fleet-wide emergency halt stops every governed agent, returns a confirmed halt count, and survives an infrastructure restart. It works for one agent or a thousand, and every halt is recorded in the audit trail.
- Where it lives
- Halt operation in the API reference · Core concepts: emergency halt
- How it is proven
- GovernanceBench override suite · Published result manifest
- Where it stops
- A halt reaches every governed agent. An agent that has not joined the layer is outside its reach, which is why discovery comes first.
How do I prove AI agent compliance for the EU AI Act, SOC 2, or HIPAA?
Every governance decision is written into a tamper-evident, hash-linked audit chain you can export and independently verify. The platform maps controls to the EU AI Act, SOC 2, HIPAA, ISO 27001, and PCI DSS. These are self-assessed readiness mappings, not third-party certification, a distinction we state plainly, because a governance product cannot claim what it has not earned.
- How it is proven
- GovernanceBench audit suite · Published result manifest
- Where it stops
- The framework mappings are self-assessed readiness mappings, not third-party certification. Agentomy does not certify you; it gives you the evidence to be certified.
How do I test my agent platform against the OWASP Agentic Top 10?
Point the open GovernanceBench and Agentomy VIGIL suites at any platform, including your own. GovernanceBench scores six governance dimensions, VIGIL runs an adversarial and prompt-injection battery, and the Kevlar suite covers the OWASP Agentic Top 10. The score is honest because the methodology is open and the benchmarks are reproducible.
- Where it lives
- GovernanceBench · Developer docs
- Where it stops
- A benchmark score describes the scenarios it ran. The manifest names the runner versions and scenario sets behind each figure, so the scope of a score is stated, never implied.
Does my data leave my environment?
It depends on the deployment mode, and the answer is stated per mode rather than universally. Self-hosted: governance runs where your agents run and your data never leaves your environment. Shared and dedicated plans run on Agentomy-operated infrastructure, so governance decisions and audit records for those workspaces are processed and retained there for the plan's retention period. Dedicated (Enterprise) is an isolated per-customer instance operated by Agentomy. In every mode you bring your own model key; on Agentomy-hosted plans it is encrypted at rest, held only as ciphertext with a masked hint, and used solely to call your model.
- Where it lives
- Pricing: plans and deployment modes · Machine-readable plans
- How it is proven
- Adapter profiles: what each adapter sends and never sends
- Where it stops
- The answer is per deployment mode and is stated on the pricing page. On a shared or dedicated plan, governance decisions and audit records are processed on Agentomy-operated infrastructure for the plan's retention period.
Which agent frameworks does Agentomy support?
Agentomy is vendor-neutral by design. It governs agents across frameworks, models, and clouds through framework adapters and standard integration paths (SDK, REST API, MCP, lifecycle hooks, CLI), so no auditor is ever asked to accept the agent vendor grading its own homework.
- Where it lives
- Install: integration paths · Adapter capability profiles
- How it is proven
- Pattern provenance: patterns per family and suite
- Where it stops
- Governance coverage is not integration coverage. An adapter brings an agent into the layer and governs a specific set of operations in a specific way; its profile says which.