AWS Deep Insight: four context layers cut script tokens about 95%

On 4 November 2026 the AWS Machine Learning Blog published Part 2 of a series on Deep Insight, an open-source multi-agent data-analysis project built on the Strands Agents SDK and Amazon Bedrock AgentCore. The post, by AWS authors, sets out four layers of context engineering and says that applied together they make "a measurable difference" on long agent tasks.

The problem the post describes

AWS uses a request such as "Analyze sales data and create a DOCX report". It says that runs through five stages (data loading, category analysis, five charts, statistical validation, report) taking about 15 minutes and over 150,000 tokens, against the 200,000-token window of Claude Sonnet 4.5 on Amazon Bedrock. The post lists the failure modes: the window overflows mid-task, stale context degrades output, and a mistake forces a full re-run.

The four layers

• Architecture: each specialist agent (Coder, Validator, Reporter and others) gets its own context and passes back only a compressed summary. AWS says this keeps even the busiest agent, the Coder, under 25,000 tokens in total.
• Prompts: each agent has an output budget, which the post gives as 1,000 to 1,500 tokens for the Coder, 800 for the Validator and 1,000 for the Reporter. Scripts must be self-contained, so old code can be dropped from context.
• Files, not messages: a single write-and-execute tool saves a script, runs it and returns only a short summary and the first 500 characters of output. Detailed results go to files that other agents read when needed. Specialist guides load as skills on demand.
• Verification: a Validator re-runs the Coder's calculations and flags mismatches of 0.01 or more, and an Auditor checks the finished DOCX against the verified citations and can request up to two Reporter retries. A summarising conversation manager is the last safety net.

Cost controls in the design

The post says prompt caching is switched on per agent, only where the same prompt is called repeatedly: the Supervisor, Coder, Reporter and Auditor have it, while the one-shot Coordinator and Planner and the low-hit-rate Validator and Tracker do not. It states that cached input tokens can cost up to 90% less on repeated calls. To limit checking cost, the Validator re-runs every high-importance calculation but only a sample of medium and low ones: above 50 calculations it checks all high and 10 medium.

What the post leaves out

The 95% and 97% figures are the authors' own worked examples (about 100 lines of code against about three lines of summary; about 20,000 tokens against about 500), not benchmarks of end-to-end runs. The post reports no dollar cost per report, no error rate with and without the Validator and Auditor, and no latency. Because it is written by the vendor to promote its own SDK and services, treat the architecture as a checklist to test, not as proof of savings.

The transferable part does not depend on AWS: keep generated code and bulky results in files, give each agent a word budget, load specialist instructions only when called for, and have a separate step recompute any number a report will quote.