Contextual Security-Aware Architecture
A single map for governed context, agent proposals, independent authority, business commits and evidence across the enterprise.
Written as at 22 October 2026.
agent-securityenterprise-architecturecontext-engineeringzero-trustassurance
An enterprise agent can read a correct document and still take the wrong action. The document may belong to another tenant. The employee who asked may lack authority for this transaction. The agent may call an approved tool with an unapproved payload. A successful tool response may only mean that work was queued, not that the business record changed. A security architecture has to join these facts at the moment of action.
I call that joined view the Contextual Security-Aware Architecture. CAE supplies governed context for decisions. The security architecture checks identity, delegation, tool scope and consequence outside the model, then records what the authoritative system actually did. It is a way to see where controls meet; it is not a new product or a claim that a model can authorise itself.
One request across five boundaries
Imagine MAAH Group, the book’s illustrative manufacturer. A procurement agent sees a supplier-risk alert and proposes holding a purchase order. The same workflow crosses business data, a model, an agent runtime, an ERP gateway and a system of record. Each crossing answers a different question.
| Boundary | Question that must be answered | Independent evidence |
|---|---|---|
| Context | Is this source current, classified correctly and available to this task and tenant? | Source, entitlement and Context Contract version |
| Delegation | Which named human and agent may propose this class of action now? | Workload identity, owner and narrowed authority |
| Execution | Is the requested tool, payload and destination within the allowed schema and budget? | Policy decision and exact command digest |
| Commit | Has the required approval been bound to this command, and will the ERP accept it? | Approval, transaction ID and ERP outcome |
| Evidence | Can a separate reviewer reconstruct the decision without trusting the agent’s account? | Protected decision and outcome record |
The model can help interpret a supplier notice, identify related orders and draft a proposal. It cannot grant itself access to the supplier master or decide that its own explanation counts as approval. A deterministic gate evaluates the exact proposed command. The ERP still validates and commits the business change. If the ERP refuses it, the workflow reports refusal even if the agent’s tool call returned HTTP 200.
What the architecture makes visible
The same map applies across IT, cloud, R&D, supplier and operational-technology zones, but the consequences differ. Reading a design repository is not equivalent to changing a production recipe. The operating technology path needs its own independent safety layer and explicit limits on agent reach. A central control plane can register agents and distribute policy, while enforcement remains at the receiving boundary; a management dashboard alone cannot stop an unmediated path to a controller or ledger.
For a design review, trace one consequential request from source to record. Mark every point where untrusted content becomes a proposal, where authority is checked, where a person must approve, where an external system commits and where evidence is kept. Then ask the team to remove the model-based guardrail and repeat the trace. If the agent can still reach the action through a side path, the boundary is incomplete.
The book’s Figure 2.4 gives the enterprise map, and Figure 4B.1 shows the SAE assurance path within it. See the books page for the edition’s status.
References
- The Contextual Secure Agentic Enterprise, first edition, Chapters 2 and 4B (prepublication manuscript, 22 October 2026).
Related
An Agent Must Be Able to Retire
Agent retirement is a security operation: revoke identity, remove tool grants, quarantine memory and preserve evidence before deleting the runtime.
Security Logs Need a Decision Record
Tool-call logs say what an agent attempted. An assurance record also explains authority, policy, approval, system outcome and independent verification.
A Factory Is Not an API Sandbox
Manufacturing agents can analyze and propose, but physical commands require a separate safety and operations boundary that can refuse or stop the action.