Dynamic Attentional Context Scoping: Agent-Triggered Focus Sessions for Isolated Per-Agent Steering in Multi-Agent LLM Orchestration

Researcher Nickson Patel has published a groundbreaking paper introducing Dynamic Attentional Context Scoring (DACS), a new framework designed to solve the critical problem of 'context pollution' in multi-agent AI systems. When multiple LLM agents (like Claude Haiku 4.5) operate concurrently under a single orchestrator, their individual task states, partial outputs, and questions often contaminate each other's decision-making context, severely degrading performance. DACS addresses this through an elegant two-mode architecture where the orchestrator maintains lightweight status summaries (≤200 tokens per agent) in Registry mode, then switches to Focus mode when an agent requests steering, injecting that agent's full context while compressing all others.

The system's effectiveness was rigorously tested across 200 trials in four experimental phases, examining scenarios with 3-10 concurrent agents, adversarial dependencies, and high decision density. Results were dramatic: DACS achieved 90.0-98.4% steering accuracy compared to just 21.0-60.0% for traditional flat-context approaches, with statistical significance (p < 0.0001). Cross-agent contamination—where one agent's context interferes with another's—plummeted from 28-57% to 0-14%. The accuracy advantage grew with both the number of agents (N) and decision density (D), showing the system scales effectively. Context efficiency ratios reached up to 3.53x, meaning DACS delivers substantially more accurate steering while using context windows more intelligently.

This represents a fundamental advance in AI orchestration architecture. Rather than relying on complex compression or retrieval techniques, DACS uses deterministic, agent-triggered context isolation. When an agent emits a SteeringRequest, the orchestrator dynamically scopes attention to that specific agent's full context while maintaining only minimal registry entries for others. This asymmetric approach eliminates the need for all agents to compete for the same context window space simultaneously. The validation included LLM-as-judge assessments across all phases with high agreement (mean kappa=0.909), confirming the robustness of the keyword matching and decision quality improvements.