01 What happened
On November 4, 2026, AWS published a blog post detailing four layers of context engineering for production AI agents, using the open-source Deep Insight reference implementation, built on the Strands Agents SDK and Amazon Bedrock AgentCore. The post quantifies token reductions: a combined write-and-execute tool saves ~95% of tokens per script, and loading skills on demand cuts the system prompt from over 20,000 to about 500 tokens at startup — a ~97% reduction.
02 Key details
- AWS proposes four context engineering layers: architecture isolation, prompt control, file-based externalization, and verification.
- Deep Insight's write-and-execute tool collapses file writing and execution into one step, saving ~95% of tokens per script.
- Loading skill guides on demand reduces the system prompt from over 20,000 to ~500 tokens (~97% reduction).
- Selective prompt caching for frequently called agents can cut input token costs by up to 90%.
03 Why it matters
AI agents handling complex data analysis tasks frequently overflow the 200K-token context window, causing failures and higher costs. The described techniques provide practical methods to manage context efficiently.
04 Who it matters to
AI agent developers, machine learning engineers, information security specialists, enterprise administrators and technical directors.