Structured Authority Injection: Enhancing Human Control in LLM Agents

Structured Authority Injection (SAI) is a proposed protocol designed to enhance human control over LLM-based agents, which are increasingly utilized as autonomous orchestrators in distributed and multi-agent systems. Despite their growing use, a significant control gap exists: there is currently no protocol-level mechanism that allows human operators to inject binding directives that persist across planning cycles.

One of the main challenges with existing free-text overrides is their susceptibility to ambiguous interpretation. This ambiguity arises from the lack of schema validation and the erosion of authority as operator instructions lose influence within the accumulating context of the agents. SAI addresses these issues by routing operator corrections through typed, schema-validated Model Context Protocol (MCP) tool calls, while also maintaining a ranked authority context that is prepended to each LLM prompt.

Override Classes and Authority Management

SAI defines three distinct override classes: Constraint, Priority, and Strategy. Each class has unique persistence semantics and an explicit precedence order that is enforced through a directed authority channel, which is inaccessible to the planning process. This structure ensures that operator directives are not only clear but also maintain their intended influence throughout the planning iterations.

Importantly, SAI does not require any model fine-tuning or access to model weights, making it applicable to instruction-tuned LLMs that support MCP tool use. Evaluations conducted across nine different LLMs and eight scenarios targeting various agent failure modes demonstrate that SAI significantly reduces ambiguous directive interpretation and mitigates authority decay across planning iterations.

Evaluation Results

The results indicate that SAI leads to lower prompt overhead in the strongest models, achieving reductions of up to 38% depending on the model and scenario. However, the benefits of SAI vary substantially across different models, highlighting the need for further exploration in diverse contexts.

In conclusion, SAI provides a general design pattern for agent systems that require explicit and persistent human control, addressing a critical gap in the management of LLM-based agents.

Source: arXiv cs.AI