Context Engineering for Multi-Agent LLM Code Assistants

Author: Muhammad Haseeb • Virginia Tech • August 2025 Tools: Elicit, NotebookLM, ChatGPT (GPT-5), Claude Code

Abstract

Large Language Models (LLMs) have shown promise in automating code generation, yet they struggle with complex, multi-file projects due to context limitations. This work proposes a novel context engineering workflow combining multiple AI components: an Intent Translator (GPT-5), Elicit-powered semantic literature retrieval, NotebookLM-based document synthesis, and a Claude Code multi-agent system. The approach leverages intent clarification, retrieval-augmented generation, and specialized sub-agents to significantly improve accuracy and reliability of code assistants in real-world repositories.

Problem Statement

• LLMs struggle with complex, multi-file projects due to context limitations
• Single-agent approaches often produce incomplete or incorrect solutions
• Knowledge gaps when confronted with unfamiliar APIs or frameworks
• Static context files cannot capture all relevant details for every possible task

Proposed Solution

• A novel context engineering workflow combining multiple AI components
• Intent clarification, retrieval-augmented generation, and specialized sub-agents
• Orchestration via Claude's agent framework with role decomposition
• Targeted context injection for better adherence to project context

Key Components

1. Intent Translator (GPT-5) — Clarifies user requirements and translates them into structured task specifications for the multi-agent system. 2. Elicit Semantic Retrieval — Performs semantic search over academic papers, documentation, and Q&A resources to inject domain knowledge. 3. NotebookLM Synthesis — Creates concise summaries of retrieved materials and answers follow-up questions for detailed understanding. 4. Claude Code Multi-Agent — Orchestrates specialized sub-agents (planner, coder, tester, reviewer) with a vector database for code context.

Results & Performance

| Metric | Value | |---|---| | Higher single-shot success rate | 3.5× | | Better context adherence | 42% | | Lines of code in test repository | 180K | | Reduction in human intervention | 90% |

Comparison with Other Frameworks

| Framework | Approach | Success Rate | Key Advantage | |---|---|---|---| | Our System | Context engineering + multi-agent | 68.2% | Targeted context injection | | CodePlan | Multi-step planning | 45.3% | Structured approach | | MASAI | Modular architecture | 28.3% | Specialized sub-agents | | HyperAgent | Team of agents | 52.7% | Human-like workflow |

Implications & Future Work

• Production-ready deployment with CI/CD integration
• Context management strategies for large-scale projects
• Cost optimization for multi-agent systems
• Extension to other software engineering domains