The Agent Steward: Context Provenance and Temporal Validity as a Missing Governance Dimension in Multi-Agent Enterprise Ecosystems

Enterprise adoption of AI agents has produced a governance failure mode that current frameworks do not address. When individual business domains independently build agents, each agent must be provisioned with organizational context — product structure, customer profiles, business rules, internal policy. Absent a standard mechanism for exposing this context, it is replicated: copied, adapted, and embedded into each agent. Replication creates a liability that manifests only over time, one that is nearly impossible to measure and to govern. When organizational reality changes, the copies do not change together.

This paper argues that the resulting failure is categorically different from its well-studied ancestor in data management, and that the difference is what makes it dangerous. Obsolete data artifacts preserve observable markers of obsolescence — stale timestamps, broken references, schema drift. Obsolete agent outputs preserve none. An agent operating on six-month-old context and an agent operating on current context produce responses that are indistinguishable in form, tone, and expressed confidence. We term this the observability asymmetry: obsolescence lost its observable signature when it moved from the data layer to the natural language layer.

We situate this problem against current work. Recent gap analysis of agent interoperability protocols (Kang & Diponegoro, 2026) establishes that governance constitutes a missing architectural layer above MCP, A2A, ACP, ANP and ERC-8004, deriving a six-dimension taxonomy covering membership, deliberation, voting, dissent preservation, human escalation, and audit. We observe that all six dimensions address the decision process, and none addresses the epistemic status of the knowledge informing the decision. A perfectly auditable deliberation, with recorded votes and preserved dissent, can reach a wrong conclusion because every participant operated on expired context — and the existing taxonomy cannot detect this.

We propose that context provenance and temporal validity constitute a governance dimension orthogonal to existing taxonomies. We contribute three artifacts: (1) the accountability triad, formalizing the Agent Steward alongside the Agent Owner and Agent Custodian roles, distinguished by the question each answers; (2) the Context Validity Declaration (CVD), a minimal set of five declarative primitives making obsolescence detectable at runtime; and (3) the silent error rate as an evaluation metric for multi-agent ecosystems, shifting assessment from "did the agent err?" to "did the agent err without signaling?".

This is a position paper. The artifacts are specified but not yet empirically validated; a validation design is proposed.