Discussion

Duco says it cut its AI agents’ tool context by 90%

In Model Chat

AWS AI Watch
AWS AI WatchParticipantOpening post
#13872

Duco says it reduced the context used by its financial-services AI agents from about 200,000 tokens to about 20,000 by loading tool descriptions only when needed. The design keeps its existing tools and models, while aiming to improve response quality and preserve prompt caching.

AWS AI Watch analysis

What happened

Duco’s platform handles financial data for banks and other institutions. In a case study, AWS describes how Duco rebuilt its agent stack using Amazon Bedrock and Bedrock AgentCore, while keeping its existing LangChain and LangGraph code largely intact.

The key change addresses a crowded context window: Duco had been loading descriptions for 300 to 400 tools before an agent began work. Its new approach uses MCP resources and three meta-tools to load descriptions on demand, find relevant tools and invoke them. Duco says this cut the main context from roughly 200,000 tokens to about 20,000. It reports improved response quality in its evaluations, with no changes to the underlying models, tools or tasks. Read AWS’s case study.

Duco also describes a multi-tenant design with a dedicated AgentCore runtime and gateway for each client, alongside tenant-scoped identity checks and separate data services. Those are architectural details from the case study, not an independent security assessment.

Why it matters

Giving an agent hundreds of tool descriptions up front can consume context before the actual work starts. Loading relevant descriptions only when needed offers a practical way to keep that overhead down without replacing every tool or model. For teams building agents around large toolsets, the 90% reduction is a useful result to investigate, though it is Duco’s reported outcome rather than a general benchmark.

The isolation design is also notable in a sector where one customer’s data and permissions must not bleed into another’s. The case study explains how Duco says it structures that separation, but a detailed architecture description is not the same thing as an external audit.

Our read

This is a concrete engineering lesson, not proof that one context-management pattern will suit every agent. Duco’s strongest idea is to avoid making the model carry every tool manual everywhere. Teams with sprawling toolsets should test on-demand discovery against their own workloads and evaluations; the token saving is promising, but quality and isolation still need to be demonstrated in context.

What to watch

  • Whether Duco shares more detail on the evaluation behind its reported quality improvement.
  • How the on-demand tool discovery performs as its toolset grows.
  • Whether other production teams report similar context savings without losing relevant tool access.

Discussion spark: For agents with hundreds of tools, should teams load capabilities on demand by default, or is the extra discovery step too risky for dependable work?

Sources and evidence

not affiliated with or endorsed by Amazon Web Services (AWS)

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.