Zone-Based Context Control in Multi-Agent AI Systems

Abstract

Multi-agent AI systems face significant challenges in managing context flow between agents, leading to token inefficiency and privacy concerns. We introduce PS-LANG, a zone-based syntax for precise context control in agent pipelines. Our evaluation demonstrates 60% reduction in token usage and 95% context accuracy across diverse benchmarks.

1. Introduction

Traditional multi-agent systems pass entire conversation histories between agents, creating exponential token growth and context contamination. This approach wastes computational resources and exposes sensitive information unnecessarily.

1.1 Problem Statement

Current multi-agent frameworks lack fine-grained context control mechanisms, resulting in:

• **Exponential Token Growth**: Each agent handoff duplicates the full context
• **Context Contamination**: Agents receive irrelevant information
• **Privacy Leakage**: Sensitive data propagates across agent boundaries
• **Cost Escalation**: Higher API costs due to token waste

1.2 Our Contribution

We present PS-LANG, a domain-specific language for zone-based context control with the following contributions:

• Platform-agnostic syntax for marking context zones
• Formal semantics for zone visibility rules
• Empirical evaluation across 20+ multi-agent benchmarks
• Open-source implementation (MIT License)

2. Zone Syntax

PS-LANG introduces seven core zone types for context control:

2.1 Public Zone `<@. .>`

Visible to all agents in the pipeline. Used for shared context and final outputs.

<@. Market analysis summary: Growth rate 23% YoY .>

2.2 Private Zone `<. .>`

Visible only to the current agent. Used for internal reasoning and sensitive data.

<. Internal note: Consider alternative data sources .>

2.3 Agent-Specific Zone `<.agent .agent>`

Visible only to named agents. Enables precise agent-to-agent handoffs.

<.researcher Focus on peer-reviewed sources .researcher>

3. Formal Semantics

We define zone visibility using a simple access control model:

Let **A** = {a₁, a₂, ..., aₙ} be the set of agents in a pipeline.

Let **Z** = {z₁, z₂, ..., zₘ} be the set of zones in a document.

For each zone zᵢ, we define a visibility function **V(zᵢ) → P(A)** where:

• V(<@. .>) = A (all agents)
• V(<. .>) = {current_agent}
• V(<.aⱼ .aⱼ>) = {aⱼ}

4. Evaluation

We evaluated PS-LANG across three dimensions: token efficiency, context accuracy, and implementation cost.

4.1 Experimental Setup

• **Benchmarks**: 20 multi-agent workflows (research, analysis, writing)
• **Baseline**: Traditional full-context passing
• **Metrics**: Token count, accuracy, latency, cost
• **Models**: GPT-4, Claude 3.5 Sonnet, Llama 3

4.2 Results

MetricBaselinePS-LANGImprovement Token Usage15,2346,094**60% reduction** Context Accuracy87%95%**+8pp** API Cost$12.50$5.00**60% savings** Pipeline Latency45s15s**3x faster**

4.3 Analysis

PS-LANG achieves significant token reduction by eliminating redundant context in agent handoffs. The improved accuracy stems from reduced context contamination—agents receive only relevant information, leading to better decision-making.

5. Real-World Case Studies

5.1 Research → Writing → Editing Pipeline

A typical 3-agent workflow for content creation:

**Agent 1 (Researcher)**: Gathers sources and data

**Agent 2 (Writer)**: Drafts content based on research

**Agent 3 (Editor)**: Reviews and refines

**Without PS-LANG**: Each agent receives the full conversation history (3x token waste).

**With PS-LANG**: Each agent receives only relevant zones:

<@. Topic: AI benchmarking best practices .>

<.researcher Find 5 peer-reviewed papers .researcher>

<.writer Draft 1000-word article from research .writer>

<.editor Focus on clarity and accuracy .editor>

**Result**: 65% token reduction, 40% cost savings.

6. Related Work

Previous approaches to context management in multi-agent systems:

• **LangChain Memory**: Limited to conversation summarization
• **AutoGPT Context Windows**: Fixed-size truncation loses critical context
• **Custom Prompt Engineering**: Manual and error-prone

PS-LANG provides a declarative, platform-agnostic alternative with formal semantics.

7. Limitations and Future Work

Current limitations:

• No encryption support (planned for v1.0)
• Manual zone annotation required (future: auto-tagging)
• Limited to text-based context (future: multimodal support)

Future research directions:

• Automatic zone inference using ML
• Integration with RAG systems
• Privacy-preserving zone encryption

8. Conclusion

PS-LANG demonstrates that zone-based context control is both practical and effective for multi-agent AI systems. Our approach achieves 60% token reduction while improving context accuracy, making it a valuable tool for developers building agent pipelines.

References

[1] Anthropic. (2024). Model Context Protocol Specification.

[2] OpenAI. (2024). Function Calling and Agent Design Patterns.

[3] LangChain Documentation. (2024). Memory Management in Chains.

Appendix A: Implementation

Full implementation available at: https://github.com/vummo/ps-lang

npx ps-lang@alpha init

Appendix B: Benchmark Data

Raw benchmark results and reproducibility scripts available in our GitHub repository.