ASO (Agent Security Operations)¶
ASO (Agent Security Operations) is AIRS applied to a single system: one model, one boundary, three layers and a circuit breaker. Everything MASO does between agents is this pattern, extended. If you run one AI system today, start here, you are also learning MASO.
Guardrails prevent. The judge detects. Humans decide. Breakers contain. On a single system those four wrap one input and one output. On a fleet, the same four extend into the space between agents. Learn them here, prove them here, then scale them.
-
Architecture Overview
The three control layers and the circuit breaker, and how a single request passes through them.
-
Core Controls
The single-agent implementation library: classify the risk, apply the layers, deselect what you don't need.
-
Single-Agent Foundations
The whole single-agent architecture on one page, with pointers into the depth.
-
Infrastructure Controls
The infrastructure layer underneath the model: identity, network, secrets, logging, and the platform patterns.
-
Fast Lane
The shortest safe path to production for a low-risk single-agent feature.
From one system to a fleet¶
The moment your single system starts handing work to another agent, you are running a fleet, and the boundaries multiply. Every ASO control has a MASO equivalent that extends it across the hand-off:
| ASO (one system) | Becomes, at fleet scale | In MASO |
|---|---|---|
| One identity, one set of permissions | Per-agent identity, least privilege across hand-offs | Identity & Access |
| Input guardrails on one prompt | Epistemic integrity between agents | Prompt, Goal & Epistemic Integrity |
| One output check | Judging every hand-off | Objective Intent |
| One circuit breaker | Kill-switch architecture for the fleet | Environment Containment |
Where next
Running more than one agent already? Move to MASO: the same layers, secured across every hand-off.