Platform
AI agent governance for systems that act
AI agent governance means knowing what each agent may do, who it acts as and what it actually did, and being able to step in. Aicura records each agent in the same register as your AI systems, then holds it to that record while it runs.
Included in Pro and Enterprise. Registering agents and the Agent report are in the free Register plan. Compare plans
What AI agent governance covers
A support agent is given one tool for one job. Weeks later it is calling three more, under a shared key nobody can trace back to it. Agent governance is the record of what an agent is allowed to do, held against what it actually does, with a way to step in. In Aicura it comes down to six things, all attached to the agent's record in the register.
- Record
- What the agent is for, what it runs on, who owns it and how far it may act on its own.
- Identity
- A credential of its own, tied to the software that runs it, so every call is attributable to the agent and not to a shared key.
- Permitted tools
- The tools and models it may use, enforced on every request at the gateway, with budgets and rate limits.
- Decision log
- What it actually did, call by call, with the outcome and whether the tool was one it declared.
- Stepping in
- Suspension at the gateway, narrowing what it may call and revoking its credential, by a person or by Guardian as far as you allow.
- Testing
- Agent evaluations run by an engine in your own environment against the agent you actually deployed.
From first record to every call it makes
Record
Know what each agent is for and who owns it
An agent has its own kind of record in the AI register, with the fields that apply only to systems that act.
- The platform it runs on, how far it may act on its own (its action boundary) and when it must hand over to a person (its escalation rule).
- Whether a kill switch is configured, and when it was last tested.
- How long its audit trail is kept, the AI systems it runs on and the agents it is linked to.
Start from the agent library to prefill the record, and accountability stays with a named owner. Accountability for an agent cannot be delegated to the agent.
Tools
List the tools it is allowed to call
The Tools tab lists every tool the agent can call, grouped as database, API, code execution, file system, web, messaging or other.
This declared list sets what the agent may do. It feeds the gateway rules that allow only those tools, the check that runs before each deploy, and the "Not declared" flag on any action that used something else.
Identity
A credential of its own, tied to where it runs
Issue the agent a runtime identity, a credential of its own tied to the software that runs it, named by runtime, environment and workload reference. Regenerate or revoke it from the Identities tab.
- Every call the agent makes is attributable to the agent, not to a shared credential nobody can unpick later.
- Budgets and rate limits can count against each runtime identity the agent presents.
- Guardian can revoke an identity, or refuse any runtime identity the agent has not declared.
Limits
Limits checked on every request
What the agent may do is enforced on each request by guardrail rules running on a gateway in your network, the point all its AI traffic passes through.
- Approved tools. Refuse the whole request, or strip the tool and forward the rest. Tool calls in responses are always enforced.
- Approved models. Fence which models the agent may reach, including any backup models the gateway falls back to.
- Budgets and rate limits. Token or request caps per period and per minute, counted on the gateway.
- Checks before a change. A release pipeline can ask whether a new tool may be enabled before it gives the agent one.
Decision log
See every action it took
Each action the agent takes is recorded in your environment and reaches Aicura through a collector, a small component you run in your own network. The agent's Decisions tab shows them as tiles and a table.
- Decisions, blocked and completed, with the share of each.
- When, what the agent did, the outcome, the tool it called and whether the gateway allowed or blocked it.
- Filters for completed, failed, blocked, escalated and deferred decisions.
A tool that is not on the declared list carries a "Not declared" chip. Decision records are snapped into the Evidence Vault, Aicura's store of dated snapshots, as they arrive.
Stepping in
Stop an agent without cutting people off
Suspend an agent at a gateway and every tool call it makes is denied until someone resumes it. Chat traffic carries on, so the people relying on it are not cut off while you work out what went wrong.
The register records whether the agent has its own kill switch and when it was last tested, so that question has a dated answer. You can also narrow what it may call without taking it offline. Agent misbehaviour has its own incident category, linked to the agent's entry.
Stepping in
Let Guardian step in when an agent misbehaves
Guardian, Aicura's automatic response, watches figures about each agent, such as the share of calls to undeclared tools, repeated blocks on one tool and tool selection accuracy against a baseline. You choose how far it may go for each condition.
- Off, Tell me, Act with approval, Act and notify, or Just do it.
- Responses include narrowing what the agent may call, revoking its runtime identity, refusing runtime identities it never declared, putting it under review, quarantining it and re-running the risk assessment.
- Act with approval holds every response until a named approver signs off.
Testing
Test the agent you actually run
Agent evaluations run on a security testing engine inside your environment. The engine keeps its own list of what to test, matched to each agent by its internal reference, so a run hits the real deployment. The addresses and credentials it tests against stay on the engine and never reach Aicura.
Set how often it runs and when it first runs, and the engine keeps its own clock. Runs land on the agent's Testing tab with their findings, passes, failures and errors.
Questions
What is AI agent governance?
AI agent governance is how an organisation controls AI that takes actions rather than only producing output. It covers six things. A record of each agent and its owner, a credential of its own for the agent, the tools and models it may use, a log of the decisions it makes, a way to step in or stop it, and testing of how it behaves. Aicura holds all six against the agent's record in its AI register.
Why do agents need different governance from other AI systems?
Because agents act. A model that drafts a letter leaves a person between the output and the consequence. An agent plans steps, calls tools and changes other systems with less of a person in between. Its governance has to cover what it may do and who it acts as, and it has to be enforced while the agent runs, not only reviewed before launch.
What is an agent runtime identity?
A runtime identity is a credential issued to one agent and tied to the software that runs it, named by runtime, environment and workload reference. The agent presents it on every call, so the gateway can attribute traffic to that agent, count budgets against it, and refuse calls when it is revoked. Aicura issues, regenerates and revokes runtime identities from the agent's Identities tab.
What does an agent decision log record?
Each decision record shows when the agent acted, what it did, the outcome, the tool it called and whether the gateway allowed or blocked it. A tool that is not on the agent's declared list is flagged as not declared. Records arrive through a collector you run in your own network and are snapped into the Evidence Vault, Aicura's store of dated snapshots.
Can Aicura stop an AI agent, or only record what it did?
It can stop it. Suspending an agent at a gateway denies every tool call it makes until someone resumes it. Approved tool lists on the gateway refuse or strip tools it may not use, and rate limits and budgets cap its traffic. Guardian, Aicura's automatic response, can narrow what the agent may call, revoke its runtime identity or quarantine it, as far as you allow, from raising it with a person to acting on its own.
Govern your agents while they run
Agents can be registered on the free Register plan. Contact us about agent credentials, decision logs and enforcement on your own gateway.