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

Nickson Patel

cs.MAApr 9, 2026

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

Authors: Nickson Patel

Organizations: Independent Researcher

Abstract

Multi-agent LLM orchestration systems suffer from context pollution: when N concurrent agents compete for the orchestrator's context window, each agent's task state, partial outputs, and pending questions contaminate the steering interactions of every other agent, degrading decision quality. We introduce Dynamic Attentional Context Scoping (DACS), a mechanism in which the orchestrator operates in two asymmetric modes. In Registry mode it holds only lightweight per-agent status summaries (<=200 tokens each), remaining responsive to all agents and the user. When an agent emits a SteeringRequest, the orchestrator enters Focus(a_i) mode, injecting the full context of agent a_i while compressing all other agents to their registry entries. Context isolation is agent-triggered, asymmetric, and deterministic: the context window contains exactly F(a_i) + R_{-i} during steering, eliminating cross-agent contamination without requiring context compression or retrieval. We evaluate DACS across four experimental phases totalling 200 trials: Phase 1 tests N in {3,5,10} (60 trials); Phase 2 tests agent heterogeneity and adversarial dependencies (60 trials); Phase 3 tests decision density up to D=15 (40 trials); Phase 4 uses autonomous LLM agents for free-form questions (40 trials, Claude Haiku 4.5). Across all 8 synthetic scenarios, DACS achieves 90.0--98.4% steering accuracy versus 21.0--60.0% for a flat-context baseline (p < 0.0001 throughout), with wrong-agent contamination falling from 28--57% to 0--14% and context efficiency ratios of up to 3.53x. The accuracy advantage grows with N and D; keyword matching is validated by LLM-as-judge across all phases (mean kappa=0.909). DACS outperforms the flat-context baseline by +17.2pp at N=3 (p=0.0023) and +20.4pp at N=5 (p=0.0008) in Phase 4, with the advantage growing with N confirmed by two independent judges.

Figures & tables

System	$N$ concurrent	Asym. mode switch	Agent-triggered	Contamination metric
AFM	×	×	×	×
ACE	×	×	×	×
AOI	✓	×	×	×
AgentOrchestra	✓	×	×	×
AdaptOrch	✓	×	×	×
CodeDelegator	×	×	×	×

Table 1: Comparison of context management mechanisms.

$N$	Cond.	Accuracy	Contamination	Ctx (tok)	$\Delta$ acc	Ctx ratio
3	DACS	$96.7\pm 2.1\%$	$3.3\pm 3.7\%$	$561\pm 1$	$+36.7$ pp	$2.12\times$
3	Baseline	$60.0\pm 11.5\%$	$56.7\pm 13.3\%$	$1{,}191\pm 0$	$+36.7$ pp	$2.12\times$
5	DACS	$96.7\pm 5.7\%$	$14.0\pm 10.6\%$	$633\pm 25$	$+58.0$ pp	$2.72\times$
5	Baseline	$38.7\pm 10.3\%$	$52.0\pm 14.3\%$	$1{,}720\pm 151$	$+58.0$ pp	$2.72\times$
10	DACS	$90.0\pm 3.0\%$	$3.7\pm 5.0\%$	$816\pm 11$	$+69.0$ pp	$3.53\times$
10	Baseline	$21.0\pm 5.0\%$	$29.3\pm 8.2\%$	$2{,}883\pm 269$	$+69.0$ pp	$3.53\times$

Table 2: Phase 1 results: steering accuracy, contamination, and context size across

N\in\{3,5,10\}

. Values are mean

\pm

SE over 10 trials.

\Delta

: DACS minus baseline. Ctx ratio: baseline mean / DACS mean.

Figure 1: DACS vs. flat-context baseline across $N\in\{3,5,10\}$ (Phase 1). Error bars: $\pm 1$ SE (10 trials each). (a) Steering accuracy. (b) Wrong-agent contamination. (c) Average context tokens at steering time.

Scenario	Cond.	Accuracy	Contamination	Ctx (tok)	$\Delta$ acc	Ctx ratio
s4 homogeneous ( $N=3$ )	DACS	$90.2\pm 3.5\%$	$0.8\pm 0.7\%$	$815\pm 10$	$+37.7$ pp	$2.29\times$
s4 homogeneous ( $N=3$ )	Baseline	$52.5\pm 1.7\%$	$44.2\pm 3.1\%$	$1{,}869\pm 49$	$+37.7$ pp	$2.29\times$
s5 crossfire ( $N=5$ )	DACS	$96.0\pm 0.9\%$	$0.0\pm 0.0\%$	$911\pm 6$	$+59.0$ pp	$2.90\times$
s5 crossfire ( $N=5$ )	Baseline	$37.0\pm 1.8\%$	$53.0\pm 2.8\%$	$2{,}643\pm 33$	$+59.0$ pp	$2.90\times$
s6 cascade ( $N=5$ , adversarial)	DACS	$94.0\pm 1.5\%$	$7.3\pm 2.2\%$	$705\pm 10$	$+37.3$ pp	$2.65\times$
s6 cascade ( $N=5$ , adversarial)	Baseline	$56.7\pm 3.2\%$	$28.7\pm 3.4\%$	$1{,}870\pm 38$	$+37.3$ pp	$2.65\times$

Table 3: Phase 2 results across homogeneous, crossfire, and cascade scenarios. Each cell: mean

\pm

SE over 10 trials.

Figure 2: Phase 2 (agent diversity) results across s4 homogeneous, s5 crossfire, and s6 cascade. Error bars: $\pm 1$ SE (10 trials each). (a) Steering accuracy. (b) Wrong-agent contamination. (c) Average context tokens at steering time.

Scenario	Cond.	Accuracy	Contamination	Ctx (tok)	$\Delta$ acc	Ctx ratio
s7 ( $N=5,\;D=8$ )	DACS	$94.0\pm 0.8\%$	$0.2\pm 0.2\%$	$1{,}654\pm 20$	$+59.2$ pp	$3.24\times$
s7 ( $N=5,\;D=8$ )	Baseline	$34.8\pm 1.8\%$	$49.8\pm 3.3\%$	$5{,}364\pm 98$	$+59.2$ pp	$3.24\times$
s8 ( $N=3,\;D=15$ )	DACS	$98.4\pm 0.5\%$	$0.9\pm 0.9\%$	$2{,}755\pm 29$	$+54.2$ pp	$2.39\times$
s8 ( $N=3,\;D=15$ )	Baseline	$44.2\pm 1.3\%$	$51.6\pm 2.4\%$	$6{,}573\pm 153$	$+54.2$ pp	$2.39\times$

Table 4: Phase 3 results: high-density decision scenarios.

D

= decisions per agent per trial. Values are mean

\pm

SE over 10 trials.

Figure 3: Phase 3 (decision density scaling) results for s7 ( $N=5,D=8$ ) and s8 ( $N=3,D=15$ ). Error bars: $\pm 1$ SE (10 trials each). (a) Steering accuracy. (b) Wrong-agent contamination. (c) Average context tokens at steering time.

Phase	Scenario	$N$	$D$	DACS acc	Baseline acc	$\Delta$	Ctx ratio
1	s1_n3	3	$\approx 3$	96.7%	60.0%	$+36.7$ pp	$2.12\times$
2	s4_n3_homog.	3	4	90.2%	52.5%	$+37.7$ pp	$2.29\times$
3	s8_n3_dense_d3	3	15	98.4%	44.2%	$\mathbf{+54.2}$ pp	$2.39\times$
1	s2_n5	5	$\approx 3$	96.7%	38.7%	$+58.0$ pp	$2.72\times$
2	s5_n5_crossfire	5	4	96.0%	37.0%	$+59.0$ pp	$2.90\times$
3	s7_n5_dense_d2	5	8	94.0%	34.8%	$+59.2$ pp	$\mathbf{3.24\times}$

Table 5: DACS advantage as a function of decisions per agent

D

, holding

N

fixed. Accuracy delta and context efficiency ratio grow as

D

increases for the baseline; DACS accuracy is stable.

Phase	Scenario	$N$	$D$	DACS	Baseline	$\Delta$
1	s1_n3	3	$\approx$ 3	96.7%	60.0%	$+36.7$ pp
1	s2_n5	5	$\approx$ 3	96.7%	38.7%	$+58.0$ pp
1	s3_n10	10	$\approx$ 3	90.0%	21.0%	$+69.0$ pp
2	s4_n3_homogeneous	3	4	90.2%	52.5%	$+37.7$ pp
2	s5_n5_crossfire	5	4	96.0%	37.0%	$+59.0$ pp
2	s6_n5_cascade	5	3	94.0%	56.7%	$+37.3$ pp

Table 6: All 8 scenarios across three phases. DACS accuracy is consistently high (90.0–98.4%); baseline collapses under agent count and decision density pressure.

Scenario	$N$	M1 Haiku Judge		M1 GPT-4o-mini		M3 Avg Ctx Tok		$p$
		DACS	Base	DACS	Base	DACS	Base
Phase 4 — real agents (LLMAgent, Claude Haiku 4.5, 10 trials each)
ra1_n3	3	79.8%	62.6%	85.4%	67.7%	654	1,361	0.0023 **
ra2_n5	5	83.7%	63.3%	89.7%	68.2%	799	2,275	0.0008 ***
Phase 1 synthetic stubs (for comparison; values from Table 2 )
s1_n3	3	96.7%	60.0%	—	—	561	1,191	$<$ 0.0001 ***

Table 7: Phase 4 results: real-agent

N

-scaling validation. Agents are autonomous LLMs (Claude Haiku 4.5) generating free-form steering questions. Two independent judges shown. Phase 1 rows reproduce Table 2 values for cross-paradigm comparison. Welch’s

t

-test significance on per-trial Haiku-judge accuracy.

Figure 4: Phase 4 real-agent vs. Phase 1 synthetic results at matched $N\in\{3,5\}$ . (a) M1 steering accuracy (Haiku judge) for DACS and baseline across both paradigms. (b) M3 average context tokens at steering time. Error bars: $\pm 1$ SE (10 trials per condition). DACS advantage direction and growth with $N$ are consistent across paradigms; absolute accuracy is lower under real agents due to the harder free-form question regime.

Multi-Agent LLM Systems

Multi-agent large language model systems coordinate multiple AI agents to solve tasks through communication, specialization, shared tools, and memory.

See all 1,040 papers

May 28, 2026

Enhancing Multi-Agent Communication through Attention Steering with Context Relevance

LLM-based multi-agent systems have demonstrated remarkable performance on complex tasks through collaborative reasoning. However, these systems tend to rapidly accumulate extremely long conversation histories during interaction. As conversations lengthen, relevant information is increasingly diluted by irrelevant context, leading to degraded performance. In this work, we present Agent-Radar, a training-free context management method that dynamically steers each agent's attention toward relevant context with a novel temporal and spatial decay mechanism. Our experiments demonstrate that Agent-Radar outperforms state-of-the-art methods across five different benchmarks, yielding gains of up to 7.64 absolute points. Furthermore, our analysis shows that Agent-Radar remains effective and robust as the number of agents and interaction rounds increases. Finally, the ablation study shows that core components in Agent-Radar are crucial to performance and generalizable in different settings.

Hongxiang Zhang, Yuan Tian, Tianyi Zhang

Purdue University West Lafayette, IN, USA

Multi-Agent LLM Systems Multi-Agent System Optimization Multi-Agent Communication LLM Context Management

Ask PDF

Aug 6, 2025

RCR-Router: Efficient Role-Aware Context Routing for Multi-Agent LLM Systems with Structured Memory

Large language model (LLM) agents are increasingly organized into multi-agent systems for complex reasoning and decision-making tasks. However, scaling such systems can substantially increase inference cost, especially when multiple agents repeatedly process overlapping or role-irrelevant context. Many existing multi-agent frameworks focus on orchestration and role specialization, but still rely on full-context sharing or static context routing, leading to redundant token consumption, increased runtime, and inefficient memory utilization. We present Role-Conditioned Context Routing (RCR), an execution-time layer that allocates role-relevant shared-memory items to agents under token budgets. Given a structured shared memory and a set of role-specialized agents, RCR selectively routes task-relevant memory items to each agent according to its role, optional task-stage metadata, and token budget. RCR integrates role-conditioned relevance scoring, token-budgeted memory selection, and iterative memory updates to reduce redundant context exposure during multi-agent execution. Experiments on multi-hop QA, embodied reasoning (ALFWorld), and web-based decision-making (WebShop) show that RCR reduces token usage by 35% on average and runtime by 38% on average relative to Full-Context, while maintaining or improving task performance.

Jun Liu, Zhenglun Kong, Changdi Yang +12

Carnegie Mellon University · Northeastern University · Harvard University +5

LLM Context Management Multi-Agent LLM Systems Multi-Agent Systems Shared-Memory Multi-Agent Systems

Ask PDF

May 29, 2026

Learning Agent-Compatible Context Management for Long-Horizon Tasks

LLM agents increasingly face long-horizon tasks such as web search and deep research in real-world applications, where accumulated context can cause long-context degradation and reasoning failures. Prior work mitigates this through context management with agent-side context control or fixed strategies such as summarization, which require training the agent itself for adaptation - making it impractical for closed-source agents and ignoring that different agents may require different strategies. We introduce Adaptive Context Management (AdaCoM), which trains an external LLM to manage the context of a frozen agent through flexible modification actions and end-to-end reinforcement learning. Across diverse agents on web search and deep research benchmarks, AdaCoM substantially improves performance by preserving task constraints and progress while pruning stale content. The learned strategies reveal a Fidelity-Reliability Trade-off: agents with higher vanilla ReAct performance benefit from higher-fidelity context preservation, whereas lower-performing agents require more aggressive compression to stay within a reliable reasoning regime. Transfer experiments show that AdaCoM generalizes most effectively across agents with similar capability (measured by vanilla ReAct performance), suggesting a practical path toward reusable context managers for agent systems.

Lu Yi, Runlin Lei, Liuyi Yao +6

1Renmin University of China · Work done during internship at Tongyi Lab, Alibaba Group · 2Tongyi Lab, Alibaba Group +2

Long-Horizon Agent Tasks LLM Context Management RL for LLM Agents

Ask PDF

Abstract

Multi-Agent LLM Systems

Explore similar work

Enhancing Multi-Agent Communication through Attention Steering with Context Relevance

RCR-Router: Efficient Role-Aware Context Routing for Multi-Agent LLM Systems with Structured Memory

Learning Agent-Compatible Context Management for Long-Horizon Tasks