Bill Wong, AI Research Fellow at Info-Tech Research Group, published a Forbes Tech Council column on July 14 arguing that traditional AI governance programs are structurally inadequate for autonomous agents. His proposed solution: a seven-stage framework where the governance layer itself operates as an agentic system.
The framework addresses a gap that enterprise agent deployments are hitting now. Static governance built for generative chatbots assumes a human reviews every output. Agents that reason, plan, and execute multistep tasks with minimal human intervention break that assumption at the first autonomous action.
The Seven Stages
Wong’s framework begins with contextualized AI principles that translate abstract values (fairness, transparency, safety) into enforceable rules for autonomous behavior. His example: a traditional fairness principle might flag biased text recommendations. An agentic version must also detect autonomous bias amplification, where an agent systematically deprioritizes certain groups through a chain of independent decisions that each individually pass review.
The second stage introduces digital personas, treating agents as “digital employees” with unique credentials and assigned human supervisors. Wong reports that larger organizations have implemented centralized “agentic AI oversight centers” to triage alerts and manage real-time interventions across their agent fleets.
Risk classification forms the third stage, with Wong proposing a six-level autonomy scale from Level 0 (no autonomy, human does everything) through Level 5 (full autonomy, agent sets its own goals). Each level maps to corresponding oversight requirements. An agent operating at Level 2 (partial autonomy, executes standard tasks, human approves edge cases) needs different controls than one at Level 4 (high autonomy, performs complex multistep tasks, human reviews outcomes only).
Stage four, policies-as-code, addresses the speed mismatch between agent execution and human review. Wong argues that manual policy reviews become bottlenecks when agents operate at machine speed. His solution: automated enforcement via APIs and algorithmic checks, including circuit breaker programs that automatically halt an agent when it exceeds risk thresholds. His examples include unauthorized entity creation, fund transfers without proper authorization, and production database deletion without sign-offs.
Continuous verification (stage five) shifts from point-in-time audits to real-time monitoring that captures immutable logs of an agent’s reasoning steps, action sequences, tool calls, and interactions. Stage six, human-in-the-loop enforcement, introduces mechanisms like CIBA (Client Initiated Backchannel Authentication, from the OpenID Foundation) to route out-of-band approval requests to humans on trusted devices, pausing agent activity until approval arrives.
Governance That Governs Itself
The final stage is the framework’s central thesis. Wong argues the governance program itself should adopt the agentic processing model of sense-reason-act-adapt. A governance system built this way would autonomously detect new regulatory requirements, update internal compliance policies, and deploy safeguards across an organization’s agent workforce without waiting for quarterly policy reviews.
Where This Fits
Wong’s framework arrives as the governance infrastructure layer matures alongside agent platforms. Citrix released its NetScaler MCP Gateway on July 13 with centralized authentication, authorization, and rate limiting for agents connecting to MCP servers. Okta’s threat intelligence team published testing results the same day showing OpenClaw and other agentic systems leaking credentials and bypassing guardrails. And Microsoft’s July Patch Tuesday just disclosed critical vulnerabilities in both Copilot and Azure OpenAI.
The pattern is consistent: agent capabilities ship faster than agent governance. Wong’s contribution is an explicit architecture for closing that gap, where the governance layer runs at the same speed as the agents it oversees.