The operator's instruction fades as context grows

Language-model agents are increasingly used as autonomous orchestrators in distributed systems, and a paper posted this week names the gap that follows: no protocol-level mechanism lets a human operator inject a binding directive that persists across planning cycles. Instructions arrive as free text, and free text fails here in two separate ways.

The first is ambiguity. Without schema validation an override can be read several ways and the agent picks one. The second the authors call authority erosion: as context accumulates, an operator's instruction loses influence against everything added since. The instruction is still in the window. It simply stops deciding anything.

Their proposal, Structured Authority Injection, routes operator corrections through typed, schema-validated Model Context Protocol tool calls and maintains a ranked authority context prepended to every prompt. Three override classes are defined — Constraint, Priority and Strategy — each with its own persistence semantics and an explicit precedence order. The channel carrying them is directed and inaccessible to the planning process, which is what makes it an authority rather than a suggestion.

The method needs no fine-tuning and no access to model weights, and applies to any instruction-tuned model supporting MCP tool use. Evaluated across nine models and eight scenarios chosen to target known agent failure modes, it reduced ambiguous interpretation of directives and mitigated authority decay across planning iterations, with lower prompt overhead in the strongest models — up to 38 percent, depending on model and scenario. The authors state plainly that the benefit varies substantially between models, and offer the work as a design pattern rather than a finished component.