---
type: "post"
id: "how-duco-transforms-financial-data-automation-with-amazon-1f1312"
slug: "how-duco-transforms-financial-data-automation-with-amazon-1f1312"
title: "How Duco transforms financial data automation with Amazon Bedrock AgentCore"
description: "Duco says it rebuilt its financial-operations agent stack on Amazon Bedrock and AgentCore, combining tenant-specific runtimes with on-demand MCP tool descriptions."
retrieval_nugget: "Duco says it rebuilt its financial-operations agent stack on Amazon Bedrock and AgentCore, combining tenant-specific runtimes with on-demand MCP tool descriptions. The design addresses two production constraints—context overload and regulated-data isolation—and AWS reports faster reconciliation deployments alongside lower onboarding costs for one bank."
published_at: "2026-11-17"
updated_at: "2026-11-17"
record_date: "2026-11-17"
date_kind: "published_at"
topics: ["mcp","claude"]
entities: ["Duco","Amazon Bedrock","Amazon Bedrock AgentCore","LangChain","LangGraph","Model Context Protocol","Anthropic Claude","Amazon Nova 2 Pro"]
related_patterns: ["agent-security-moves-to-runtime-boundaries","skills-become-portable-capability-layer"]
novelty_label: "new_fact"
what_changed: "Duco applies Amazon Bedrock AgentCore to automate financial data workflows, adding a distinct customer use case beyond prior AgentCore case studies."
prior_refs: ["https://newruntime.com/posts/how-greenlight-financial-built-an-autonomous-soc-solution-83215d/"]
source_url: "https://aws.amazon.com/blogs/machine-learning/how-duco-transforms-financial-data-automation-with-amazon-bedrock-agentcore"
source_urls: ["https://aws.amazon.com/blogs/machine-learning/how-duco-transforms-financial-data-automation-with-amazon-bedrock-agentcore"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:how-duco-transforms-financial-data-automation-with-amazon-1f1312"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/how-duco-transforms-financial-data-automation-with-amazon-1f1312/","markdown":"https://newruntime.com/posts/how-duco-transforms-financial-data-automation-with-amazon-1f1312.md","json":"https://newruntime.com/posts/how-duco-transforms-financial-data-automation-with-amazon-1f1312.json"}
markdown_url: "https://newruntime.com/posts/how-duco-transforms-financial-data-automation-with-amazon-1f1312.md"
---

# How Duco transforms financial data automation with Amazon Bedrock AgentCore

## Retrieval answer

Duco says it rebuilt its financial-operations agent stack on Amazon Bedrock and AgentCore, combining tenant-specific runtimes with on-demand MCP tool descriptions. The design addresses two production constraints—context overload and regulated-data isolation—and AWS reports faster reconciliation deployments alongside lower onboarding costs for one bank.

## Summary
Duco says it rebuilt its financial-operations agent stack on Amazon Bedrock and AgentCore, combining tenant-specific runtimes with on-demand MCP tool descriptions. The design addresses two production constraints—context overload and regulated-data isolation—and AWS reports faster reconciliation deployments alongside lower onboarding costs for one bank.

## What happened
AWS’s announcement describes Duco as an operations platform serving more than 200 financial institutions and processing 20 billion transactions monthly. The company says its earlier machine-learning systems helped automate more than 20,000 processes, but required historical training data and extensive customization. Duco adopted Amazon Bedrock for managed foundation-model access and Bedrock AgentCore as its agent runtime. Its existing LangChain and LangGraph code was largely retained, according to AWS.

The architecture separates customers throughout the stack. Each tenant has a dedicated AgentCore runtime and Gateway, its own Cognito user pool, and dedicated data and messaging services, including Aurora MySQL, S3, KMS, SES and SQS. Tenants share an EKS cluster and VPC, but use separate namespaces and network policies. Requests carry a tenant-scoped JSON Web Token; the runtime’s inbound authorizer validates it before agent logic runs. Agents call tools through the tenant’s Gateway, which connects to container-based MCP services in EKS. Those services validate both agent and user identity. AWS says each tenant’s runtime has no shared memory, context or tooling with other tenants.

Duco also uses different models for different roles: Claude Opus for planning, Claude Sonnet for complex decisions, Claude Haiku for routing, and Amazon Nova 2 Pro for summarizing tool output. The company says it can switch models through configuration. A separate design change tackles context size. With 300–400 tools exposed per tenant, static descriptions had consumed about 200,000 tokens. Duco moved descriptions into MCP resource types: `skill://` for on-demand tool descriptions, `system://` for dynamically composed prompt sections, and `ui://` for interface resources. Three harness-level meta-tools—`load_skill`, `find_tool` and `invoke_tool`—retrieve the relevant descriptions and execute selected tools. AWS reports that this reduced the main context to approximately 20,000 tokens, preserved prompt caching and improved response quality in Duco’s evaluations.

The announcement describes two workflows. For onboarding, agents interpret file formats and legacy configurations, including old spreadsheet macros, and turn them into Duco reconciliation processes. AWS says deployments are now 10 times faster. For exception investigation, agents group related breaks, check whether an issue is already known, trace evidence, recommend a resolution or escalation, and route a case to an analyst. The analyst makes the final decision, which is recorded in an audit trail. AWS reports that one global bank reduced onboarding costs by 70% and increased straight-through processing by more than 20%. Duco says the system keeps humans responsible for high-risk decisions and supports auditability and data-residency controls.

## What the archive adds
New Runtime’s earlier coverage described Greenlight Financial’s autonomous security operations center built with Amazon Bedrock AgentCore ([November 7, 2026 note](https://newruntime.com/posts/how-greenlight-financial-built-an-autonomous-soc-solution-83215d/)). That was a distinct customer deployment, not the same event. Duco adds a new fact to the broader AgentCore story: a financial-data platform using per-tenant agent runtimes and on-demand MCP skills to address context size and isolation.

## Trend signal
Duco’s per-tenant runtimes, identity checks and tool gateways support the view that production agent security depends on runtime boundaries, not prompts alone. The architecture is a concrete example of [agent security moving to runtime boundaries](https://newruntime.com/patterns/agent-security-moves-to-runtime-boundaries/). Its on-demand tool-description design also resembles a portable capability layer, though this customer account does not establish how broadly the approach generalizes.

## Why it matters
The case makes context management and authorization architectural concerns rather than prompt-tuning exercises. Loading hundreds of tool schemas into every request consumes context before the agent has begun useful work; retrieving descriptions only when needed offers a way to preserve room for the actual task. The reported reduction from roughly 200,000 to 20,000 tokens is substantial, but teams should test both task quality and latency in their own workloads.

The isolation pattern matters for teams serving multiple organizations. A dedicated runtime and gateway per tenant, coupled with identity propagation to each tool, can make access checks traceable across the full call path. That is more operationally involved than sharing one agent context, but it gives security reviewers clearer boundaries to inspect. The workflows also illustrate a useful division of labor: agents gather and structure evidence, while analysts approve consequential actions. The reported business gains are promising, but not independently validated here.

## Limits and uncertainty
The performance and business outcomes are AWS’s account of Duco’s deployments, not independently audited results. The announcement does not provide evaluation datasets, task-level quality scores, latency comparisons, model or inference costs, or details that would let another team reproduce the reported improvements. The 70% cost reduction and 20% increase in straight-through processing refer to one global bank; AWS does not describe that customer’s baseline or measurement period. The architecture is presented as a customer case, so it does not establish that the same model mix, isolation design or savings will fit other regulated workloads.

## Skill to build now
If your agent exposes many tools, try a small on-demand discovery layer: keep tool names available, retrieve a tool’s description only after the agent identifies a likely need, then invoke it through a narrow meta-tool. Compare this against the static-schema baseline on the same tasks, tracking context consumption and answer quality. If you serve multiple tenants, separately test whether identity is preserved and checked at every tool boundary.

## Reader takeaway
Duco’s most reusable lesson is to avoid choosing between a large tool surface and a bloated prompt: expose capabilities through discovery, load their details when needed, and keep tenant identity attached through execution. Treat the reported savings as a case-specific result, then validate the design against your own security and quality requirements.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Mcp](https://newruntime.com/topics/mcp.md): Explore the mcp topic hub.
- topic hub: [Claude](https://newruntime.com/topics/claude.md): Explore the claude topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)
