OpenAI announced this week the launch of Frontier, described as "an enterprise platform for building, deploying, and managing AI agents with shared context, onboarding, permissions, and governance." The timing is notable: this arrives as organizations grapple with proliferating AI agents built across multiple platforms, not just OpenAI's own tools. The company's explicit pitch is management infrastructure for heterogeneous agent ecosystems. This development surfaces a fundamental coordination problem that organizational theory has surprisingly little to say about: how do you govern autonomous systems that develop competence endogenously through interaction rather than through pre-programmed procedures?

The Documentation Fantasy in Agent Management

Frontier's feature set reveals what enterprises think they need: permissions systems, shared context repositories, onboarding workflows. These are artifacts borrowed directly from human resource management. The implicit model is that AI agents can be managed like employees if you have sufficiently detailed documentation about roles, access rights, and standard operating procedures. This represents what I have elsewhere called the proceduralization fallacy: the belief that complex coordination problems can be solved through increasingly granular specification of rules and workflows (Vergauwen, 2024).

The problem is that AI agents, particularly those involved in agentic coding or dynamic tool use as described in Anthropic's concurrent announcement of Claude Opus 4.6, do not operate through fixed procedural knowledge. They develop capabilities through interaction with their operational environment. An agent trained to write code does not follow a predetermined decision tree. It generates novel solutions based on patterns extracted from training data and refined through deployment experience. The competence is endogenous to the system, not imported from external documentation.

This mirrors the variance puzzle in platform work: workers with identical access to platform features show dramatically different performance outcomes (Kellogg et al., 2020). The difference cannot be explained by differential access to procedural knowledge, because the procedures themselves do not determine success. What matters is whether workers develop accurate structural schemas about how algorithmic systems amplify certain behaviors and dampen others. Documentation tells you what buttons to push. Schemas tell you why the buttons exist and what second-order effects they trigger.

The Governance Trap: Control Without Understanding

Frontier's positioning as a cross-platform management layer introduces a second problem. When you build governance infrastructure that sits above multiple agent systems, you necessarily abstract away from the specific operational logics of each system. You create what Rahman (2021) calls an "invisible cage": control mechanisms that constrain behavior without making the rationale for constraints transparent to either the agents or their human supervisors.

Consider permissions management for AI agents. A human resource system grants or restricts access based on role definitions that employees understand and can reason about. An AI agent operating under similar permission constraints has no comparable understanding. It encounters refusals or allowances as brute facts about its operating environment, not as expressions of organizational policy that could be questioned or negotiated. This creates what Sundar (2020) describes as machine agency without machine understanding: systems that act autonomously but cannot explain or justify their actions in terms accessible to human governance structures.

The counterintuitive implication is that more sophisticated governance platforms may actually reduce organizational understanding of agent behavior. When controls are platform-mediated and abstracted, the humans responsible for oversight lose direct visibility into why agents behave as they do. They see outputs and compliance metrics, but not the operational logics that connect governance rules to agent decisions. This is the awareness without capability problem applied to management: knowing that controls exist does not translate to understanding how those controls shape agent behavior (Gagrain et al., 2024).

What Agent Governance Actually Requires

If procedural documentation and permission systems are insufficient, what would effective agent governance look like? The answer lies in schema induction rather than rule specification. Organizations need infrastructure that makes the structural features of agent operation visible and interpretable, not just controllable. This means instrumentation that reveals how agents learn, what patterns they extract from data, and how their decision-making evolves over deployment cycles.

Frontier's "shared context" feature gestures toward this, but context sharing is not the same as schema visibility. Agents that share context pools can still develop divergent operational logics if they weight or interpret that context differently. What matters is whether human supervisors can observe and reason about those interpretive differences, not just whether agents have access to the same information.

The broader lesson is that platforms for agent management face the same transfer problem as platforms for human work: control mechanisms designed for one operational context do not automatically transfer to others. OpenAI's bet is that governance infrastructure can be abstracted and generalized across agent types. The research on algorithmic literacy suggests otherwise. Effective coordination requires structural understanding specific to each domain, not just procedural compliance enforced from above.