AIToday
Large Language ModelsAI Business & IndustryAmazon AI BlogPublished: Nov 17, 2026, 10:00 JST

Duco cuts agent context 90% on Amazon Bedrock AgentCore

Send on LINE
Duco cuts agent context 90% on Amazon Bedrock AgentCore

3 Key Points

  1. What happened

    Duco moved its agentic AI stack to Amazon Bedrock and AgentCore, reducing main context from 200,000 tokens to about 20,000 and deploying new reconciliations 10× faster.

  2. Why it matters

    Shrinking context overhead lets Duco's agents keep more of the conversation in view, which the company says measurably improved response quality validated through its evaluation framework.

WHO IT HITSEnterprise architects and platform engineering teams running multi-tenant AI agents for regulated clients are the ones who would feel this most: the body shows a way to isolate each bank's data and tools while keeping model costs and context load down.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, every morning.

Context & Analysis

Duco's platform already processed 20 billion transactions monthly for more than 200 clients, including 7 of the top 20 banks, before the rebuild. The older classical machine learning models, used for predictive labeling and pattern recognition across over 20,000 processes, required extensive historical data and struggled with novel scenarios. As Rob Gates, Duco's Chief Architect and Innovation Officer, put it, those models could only amplify what they had already learned. The move to Amazon Bedrock was a response to that ceiling, and the multi-model split by role—Opus for planning, Sonnet for reasoning, Haiku for routing, Nova 2 Pro for summaries—is designed so models can be swapped without disrupting the stack.

The context problem is the thread that ties the architecture together. With 300–400 tools exposed per tenant through AgentCore Gateway, roughly 200,000 tokens of tool descriptions were loaded before any work began, and response quality degraded as conversations lengthened. The skills architecture moves those descriptions into skill:// and system:// resources loaded on demand, which the company says cut context by 90% while preserving prompt caching because the static prefix stays small and stable. On top of that, Duco implemented MCP tool discovery at boot time, so a new microservice's capabilities reach agents without redeployment or agent reconfiguration.

The regulated-environment side is where the redesign may matter most to financial clients. Isolation is enforced through identity as well as infrastructure, with a tenant-scoped JWT validated by the AgentCore runtime authorizer and each tool call authorized as the calling user. Human-in-the-loop controls require explicit authorization for high-risk actions, and every agent decision is logged in an audit trail that reviewers can trace against regulatory requirements. Duco frames the roadmap as a shift in human roles from task worker to investigator to decision maker, with operations teams orchestrating multiple agents in parallel and making final approval calls on agent-prepared analysis.

FAQ
How does Duco keep each bank's data separate from other clients?
Each tenant gets a dedicated AgentCore runtime and Gateway configured against that tenant's own Cognito user pool and Amazon EKS namespace, so there is no shared memory, context, or tooling between clients.
Which models does Duco use for which tasks?
Anthropic Claude Opus handles planning, Claude Sonnet handles main reasoning and complex decisions, Claude Haiku handles lightweight tool routing, and Amazon Nova 2 Pro summarizes tool output.
How did Duco reduce its context usage so sharply?
Duco created a skills architecture using MCP resource types that loads tool descriptions only when needed, plus three meta-tools (load_skill, find_tool, invoke_tool) that replaced the bulky static tool list.
Amazon AI BlogRead Original Article
Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.

Questions and answers are published on this page.

Related Articles

Next articleLife insurers hold $85.2 billion in AI debt, up 24% in six months