Govern autonomous and semi-autonomous AI agents

Why agents change the risk profile

A model that produces a suggestion carries advisory risk. An agent that takes an action carries operational risk. The difference is not the model — it is autonomy, tool access, write permissions and the absence of a human between intention and effect.

Traditional model governance was never designed for that. AIRAS Cloud extends the same governed record to cover what an agent is permitted to do and what it actually did.

Agent controls in the platform

• Registered agent identity with a named accountable owner
• Declared purpose, scope of action and operating boundary
• Explicit tool allowlists and privilege boundaries
• Data and system access mapped to least privilege
• Spend, rate and volume budgets with hard ceilings
• Human approval gates for defined action classes
• Escalation thresholds and automatic hand-back to a person
• Runtime action log retained against the governed record
• Runtime moderation of prompts, retrieval, outputs and actions
• Intervention: pause, restrict scope or withdraw the agent

Evidence that an agent stayed inside its boundary

Oversight is only meaningful if it can be shown after the event. Agent actions, escalations, blocked attempts and interventions are recorded against the same append-only history as the assessment and the approval decision, so a reviewer can compare permitted behaviour with observed behaviour in one place.

Material change — a new tool, a widened permission, a raised budget, a different data source — triggers reassessment rather than quietly extending the original approval.

Aligned to recognised guidance

Agent control patterns draw on OWASP guidance for generative and agentic systems, NIST AI RMF activities and EU AI Act human-oversight themes. AIRAS Cloud operationalises those expectations; it does not issue legal opinions or certify compliance.