Which agents do we have?
The agents we supply, the ones you build and the ones you run elsewhere and register sit in one fleet list, with an inventory you can export. A caller nobody registered is flagged.
AI agent governance
Govern every AI agent from one place
Know which agents you have, see what each one did, hold risky writes for a person and stop any agent. It covers the agents we supply, the ones you build and the ones you already run.
Four questions
What a governed agent lets you answer
What did this agent do?
Follow any run step by step: each model call, tool call and guardrail check, with its tokens and timing. The run's record shows what it cost and any write held for approval.
Who approved this write?
Each write tool is allowed, held for a person or denied. An approval is tied to the exact agent, tool and arguments, so a changed request needs a new approval.
Can we stop it now?
Stop one agent, a group of agents or every agent at once. Calls already running are cut off, and the switch stays on until a person lifts it.
Limits
Limits that apply before the agent acts
Policies
Cedar policies decide which person or agent may do what. Roles apply from the start, and each policy you write narrows them.
Guardrails
Guardrails check model inputs and outputs for prompt injection, PII and denied topics.
Budgets
Spend is reserved before each model call and capped per agent, team and tenant.
The asker's own access
An agent can act with the permissions of the person who asked, never more.
Pinned tools
Each tool definition from an MCP server is pinned. If the server changes one, that tool is withheld from agents until a person accepts the change.
Autonomy ladder
Give each agent a level: shadow, draft, exceptions or autonomous. A person raises it on measured evidence, and it drops by itself when the evidence turns bad.
Evidence
Proof you can hand to an auditor
Tamper-evident audit
Every call, cost and decision in a hash-chained log you can verify and export.
Receipts you can verify
Each write an agent makes through a connected system leaves a receipt on a hash-chained ledger, tied to the approval it ran under. Export the ledger as a signed bundle and an auditor can check it offline.
A control map
The platform's controls are mapped to the EU AI Act and the NIST AI Risk Management Framework, and exported for your own assessment.
Questions
Frequently asked questions
What is AI agent governance?
It is the set of controls that let a company say which AI agents it runs, what each one may do, what each one did and who approved it, and that let a person stop any of them.
Does it work with agents we already built?
Yes. Keep agents built with any framework where they are. Register each one and send its model calls through the Liyagent gateway, and it gets an identity, guardrails, a budget, a kill switch and an audit trail.
Do we have to change our models or systems?
No. Liyagent works with the model providers you choose and reaches your systems through their own APIs, over MCP, OpenAPI or your own code.
Where does it run?
Three ways: hosted by Liyagent, in your own environment under the Enterprise licence, or as a control layer over agents you already run.
Related