PulseFlow Tecnologia
← Insights
·6 min read·framework-playbook·agent-governance·identity·audit
PulseFlow Tecnologia

A governance framework for corporate agents

When three of the world's largest cloud platforms design, at the same time, identity, registry, and audit systems specific to AI agents, that stops being vendor opinion and becomes a market signal: agent governance isn't a compliance add-on, it's part of the architecture. A practical framework can be extracted from what Microsoft, Google, and AWS actually built.

Identity and permissions

Google Cloud treats agent identity as an infrastructure problem, not a policy one: Agent Identity assigns "unique cryptographic IDs" to each agent, and Agent Registry works as a "central library of approved agents, tools, and skills" - nothing runs without being registered. Microsoft solves the same problem on top of the identity a company already uses: Entra (identity), centralized governance via Agent 365. In both cases, the logic is the same: an agent isn't an anonymous model instance - it's a traceable identity, with its own permission scope.

Data and tools

AWS approaches this dimension from the data side: Managed Knowledge Base connects agents to corporate sources (S3, SharePoint, Confluence, Google Drive, among others) through native connectors - meaning data access flows through a single, managed point, not artisanal integrations scattered across teams. On the tools side, Google Cloud uses Agent Gateway for "centralized control and unified security policies," and Model Armor specifically against prompt injection and data leakage.

Context and memory

Persistent memory has stopped being an experimental feature: Google's Memory Bank is now a formal part of the platform, with "Memory Profiles" retaining context across sessions - the Payhawk case (over 50% reduction in expense-report time) only exists because the agent remembers context between interactions. Governing this layer means deciding what gets retained, for how long, and who can audit that memory.

Logs, evals, and audit

Google Cloud has Agent Anomaly Detection (real-time suspicious behavior detection) and an Agent Security Dashboard, integrated with Security Command Center. Microsoft bets on the "source, test, deploy, observe, and improve" cycle as part of Agent 365 itself. The common pattern: none of the three companies treats observability as something added afterward - it's part of the same layer that records identity and permission.

Human approval

None of the platforms proposes full autonomy without supervision. Microsoft's continuous improvement happens "under human supervision"; Google offers Agent Simulation and Agent Evaluation before anything runs against live traffic. Mature governance isn't synonymous with blocking the agent - it's defining where human approval enters the cycle, and where the agent can act alone.

The framework's ten dimensions

Putting together what these three platforms actually implemented:

  1. Identity - a unique, traceable ID per agent
  2. Permissions - access scope defined per agent, not per whole system
  3. Data - centralized connectors, not artisanal integrations
  4. Tools - a central registry of approved tools
  5. Context - what the agent sees, and where it comes from
  6. Memory - what's retained between sessions, for how long
  7. Logs - every decision and tool call recorded
  8. Evals - tests against simulated traffic before production
  9. Human approval - defined checkpoints, not a full block
  10. Audit - continuous dashboard and anomaly detection

Sources