The choice between two incompatible models
The rapid deployment of autonomous agents has forced a strategic pivot in how enterprises manage digital risk. We are currently witnessing a divergence into two distinct governance architectures. On one side, the identity-first model, championed by Microsoft and Okta, treats agents as privileged entities within the existing security perimeter. On the other, the observability-first model, represented by MakeOS, treats agentic activity as a series of transparent, reversible processes. These models are not merely different technical implementations; they represent fundamentally incompatible philosophies regarding where control should reside in an automated workflow. The stakes of that fork rose this week, when AWS shipped the full agent stack in one keynote and left buyers to work out what they can actually turn on today.
The Identity-First Model: The Agent as Insider
The identity-first approach, detailed in Microsoft’s agentic insider framework, relies on the premise that agents should be managed like human employees. In this model, Guardian Agents act as 24/7 moderators, ensuring that every action taken by an agent aligns with predefined policies. The architecture relies on a hierarchy where people define policy, Playbooks coordinate the logic, and Guardian Agents execute the oversight. Satya Nadella has noted that the most trustworthy agentic system will not be the one with the model we trust most, but rather the one with the most robust wrapper. This strategy focuses on securing the non-human identity to prevent unauthorized access.
The Observability-First Model: The Agent as Process
Conversely, the observability-first model, outlined in MakeOS’s glass box approach, shifts the focus from the actor to the action. Instead of attempting to secure the agent’s identity, this model prioritizes visibility into the execution chain. Every action is designed to be visible, permissioned, and, crucially, reversible. Patrik Simek describes this as being business deterministic where it matters, agentic when it helps, and human when it counts. By treating the agent as a process rather than an identity, the system aims to mitigate risk through continuous auditability rather than static policy enforcement.
The Audit Burden: Efficacy vs. Visibility
Each model imposes a distinct operational burden on compliance teams. The identity-first model requires rigorous auditing of the wrapper and the translation of human policy into machine-readable constraints. If the policy translation fails, the agent operates outside of its intended scope. The observability-first model, however, requires a different set of competencies. Compliance teams must audit process logs and ensure that every automated action can be effectively reversed if it deviates from business requirements. The challenge here is not just policy definition, but the technical capability to maintain a granular, real-time audit trail of every agentic decision.
The Compliance Disconnect: Confidence vs. Control
The current state of enterprise adoption reveals a significant gap between executive perception and operational reality. According to the Okta AI Agents at Work 2026 survey, 96% of executives express confidence in their identity and access management for non-human identities. Yet, only 34% of those same organizations apply the same level of controls to those identities as they do to human users. This disconnect is compounded by the fact that 52% of employees are currently using unapproved AI tools. Consequently, 58% of executives reported an incident or a close call within the past year, highlighting the fragility of current governance frameworks.
Conclusion: Who Carries the Liability?
The choice between these models ultimately dictates where the liability rests when an automated process fails. Identity-first models place the burden on the integrity of the policy wrapper and the security of the identity itself. Observability-first models place the burden on the transparency of the process and the ability to revert actions. As organizations continue to integrate tools like Microsoft 365 E7 security packaging, the decision to prioritize identity or observability will define their resilience against shadow AI risks and the inevitable operational errors that accompany agentic workflows. Leadership must decide whether they are governing the actor or the action.