Beyond the Context Window Limit - Practical Context Engineering Techniques Learned from the Deep Insight Development Journey

Building an AI Agent to production level reveals a different reality than simple Q&A: data exploration, analysis, verification, and report generation all chain together within a single task, immediately exposing problems like Context window overflow, performance degradation, and re-execution costs. To solve this, the AWS Korea SA Team built Deep Insight, a production-grade Multi-Agent system, and organized the Context Engineering techniques validated during that process into 4 layers.

The key starting point is that even Claude Sonnet 4.5's 200K Context window can easily fall short in real work flows. For example, a task that analyzes sales data and produces a DOCX report—moving through load/exploration, category-wise sales analysis, time-series analysis and chart generation, statistical verification, and final report writing—can consume a total of 15 minutes and over 150K tokens. Putting everything into a single Context causes intermediate failures, performance degradation from stale Context, full re-execution, and increased token-based costs, all at once.

Deep Insight addresses this by dividing the problem into the following 4 layers.

• Layer 1: Architecture - Separates Context using a multi-agent structure. Coordinator, Planner, Plan Reviewer, Supervisor, Coder, Validator, Reporter, and Tracker each work in their own independent Context, and shared information is limited to only messages, clues, full_plan, and history.
• Layer 2: Prompts - Explicitly specifies output token budgets for each agent, and enforces self-contained code principles and completion rules via prompts to prevent unnecessary Context inflow. Coder is limited to around 1,000-1,500 tokens, Validator to 800 tokens, and Reporter to around 1,000 tokens.
• Layer 3: Tools - Externalizes heavy code and intermediate outputs to files instead of accumulating them in Context. The write_and_execute_tool handles code writing and execution in one step so the code body isn't left in Context, and common functions are reused via modules like coder_analysis_utils.py.
• Layer 4: Verification/Safeguards - Ensures quality using structured files such as all_results.txt, calculation_metadata.json, and citations.json along with a verification agent, serving as the final line of defense against Context overflow.

In particular, Structured Note-Taking is core to inter-agent collaboration: detailed results from each task are accumulated in all_results.txt, and the next agent reads it only when needed. As a result, the message the Coder passes to the Supervisor is reduced to something like "Analysis complete, 3 key insights found, detailed results saved to file," while lengthy analysis content is safely stored in the file system. Ultimately, what Deep Insight demonstrates is clear: Context Engineering is not a matter of a single prompt line, but rather the design of an entire system that conserves Context by jointly designing architecture, tools, and verification.