Duco transforms financial data automation by using Amazon Bedrock AgentCore.
Duco uses Amazon Bedrock and AgentCore to run isolated financial AI agents, cutting deployment time and onboarding costs for banks and asset managers.
Key Takeaways
- Duco shifted from traditional machine learning models to agentic AI on Amazon Bedrock and AgentCore to improve reasoning, adaptation, and automation of financial data processes.
- The platform supports 20 billion monthly transactions for 200+ clients, reducing deployment time and onboarding costs while serving regulated financial institutions.
- Amazon Bedrock enables access to multiple foundation models through a single API, with customer data kept inside AWS and private endpoints used to connect securely to internal systems.
- AgentCore was selected for compatibility with LangChain and LangGraph, a managed runtime with observability, and enterprise-grade security and compliance controls for production AI agents.
- Multi-tenant isolation is enforced through dedicated AWS resources, EKS namespaces, JWT and Cognito authentication, tenant-specific AgentCore Gateways, MCP tool access, and session isolation.
Meta description: Duco uses Amazon Bedrock and AgentCore to run financial AI agents, reduce deployment time, lower onboarding costs, and keep each bank's data isolated.
Tags: Amazon Bedrock, AgentCore, Duco, financial services AI, multi-tenant isolation
Featured image: Abstract friendly illustration of isolated data lanes and AI agents, no people, no logos, no text.
This article is for financial services technology teams evaluating agentic AI in regulated environments. It summarizes how Duco, an agentic Operations platform for financial services, moved its AI workloads to Amazon Bedrock and Amazon Bedrock AgentCore. The source says Duco processes 20 billion transactions monthly for 200+ clients, including 7 of the top 20 banks and 10 of the top 20 asset managers, and that the rebuild reduced deployment time and onboarding costs while keeping each bank's data isolated.
From classical models to agentic operations
Duco's platform lets teams automate mission-critical data processes using AI and no-code models, covering tasks from data extraction to downstream publication. Before the change, it relied on traditional machine learning models for predictive labeling and pattern recognition, helping banks consolidate and automate more than 20,000 processes.
Those classical models had limits. They required extensive historical training data, struggled with novel scenarios, and needed significant customization for each new use case. They also had difficulty adapting quickly to the changing data formats and business rules found in financial operations. Rob Gates, Chief Architect and Innovation Officer at Duco, said the earlier prediction models were not agile or flexible enough and could mainly amplify what they had already learned rather than identify new problems or guide the next step.
As data processing demands grew, Duco moved toward agentic AI capabilities built on Amazon Bedrock and Amazon Bedrock AgentCore. The goal was not only to add a new model, but to change how the platform could understand, reason about, and act on financial data.
Why Bedrock and AgentCore were selected
Amazon Bedrock is a fully managed service that offers access to foundation models from leading AI companies through a single API. Duco can choose from models in the Amazon Nova 2 family, Anthropic, OpenAI, Cohere, Meta, and Mistral AI, and can switch models by changing a configuration parameter. The source says customer data stays within the AWS environment and is not used to train models. Duco can also use private endpoints in its virtual private cloud (VPC) to connect to internal systems while keeping them isolated.
Amazon Bedrock AgentCore is positioned as a platform to build, connect, and optimize AI agents at scale. It supports open source frameworks, allows agents to connect to tools and data with authentication and access controls, and provides tracing and evaluation for agents in production. AgentCore handles infrastructure, so teams can focus on agent logic rather than backend plumbing while keeping enterprise security and reliability expectations.
Duco's existing agent infrastructure was built on LangChain and LangGraph, both already in production. AgentCore was selected for three main reasons: compatibility with existing LangChain and LangGraph code, a managed runtime that reduces operational overhead, and suitability for regulated financial services environments with built-in security and compliance controls.
How isolation is enforced across tenants
Duco operates as a multi-tenant software as a service (SaaS) platform with strict tenant isolation. For regulated institutions, this is treated as a non-negotiable requirement. The architecture applies isolation from the cloud infrastructure up to the AI agent runtime.
At the infrastructure layer, Amazon Elastic Kubernetes Service (Amazon EKS) clusters use namespace-per-tenant isolation and network segregation to separate workloads. Each tenant receives dedicated AWS services, including Amazon Aurora MySQL, Amazon Simple Storage Service (Amazon S3) storage, AWS Key Management Service (AWS KMS) encryption keys, Amazon Simple Email Service (Amazon SES) and Amazon Simple Queue Service (Amazon SQS) messaging, Amazon ElastiCache, and an Amazon Cognito user pool. The source states there is zero shared data infrastructure across clients. Per-tenant autoscaling helps reduce the impact of one client's compute-intensive workload on another tenant's performance.
Access control is enforced through identity as well as infrastructure. Every request carries a JSON Web Token (JWT) issued by the tenant's dedicated Cognito user pool. That token is validated by the inbound authorizer of the AgentCore runtime before agent logic runs. Agents reach tools only through the tenant's dedicated AgentCore Gateway, which forwards calls to the tenant's container-based Model Context Protocol (MCP) services running in Amazon EKS.
The user's identity is passed through to downstream services. Each MCP server validates both the agent's identity and the user's identity against the tenant's Cognito user pool, so every tool call is authorized as the calling user, made by a trusted agent.
Tenants share a VPC and Amazon EKS cluster, but traffic is segregated by namespace and network policy. Private endpoints keep traffic to Amazon Bedrock and AgentCore on the AWS network. Each tenant's AgentCore runtime runs in VPC network mode, with network interfaces in private subnets, so outbound traffic enters the VPC and is governed by security groups.
At the agent layer, each tenant has one dedicated AgentCore runtime and one dedicated Gateway, configured against that tenant's own Cognito user pool and Amazon EKS namespace. Combined with AgentCore's built-in session isolation, this means there is no shared memory, context, or tooling between clients. Each bank's schemas, naming conventions, and institutional knowledge remain isolated from other tenants. Each tenant also has its own set of MCP servers, dynamically discovered at boot time. The design supports a model where each agent can access only one bank's data, context, and reasoning.
Multi-model routing and context control
Rather than using one foundation model for everything, Duco assigns each role in its agent design to a model suited to that task. This is intended to balance quality and cost efficiency while allowing model changes without disrupting the stack.
| Model | Primary role | Use cases |
|---|---|---|
| Anthropic Claude Opus | Planning agent | Complex reasoning, understanding, and formulating plans for other agents to execute |
| Anthropic Claude Sonnet | Main reasoning and complex decision-making | Exception investigation, reconciliation logic, and rule interpretation |
| Anthropic Claude Haiku | Lightweight tool routing decisions | Fast routing choices, low-latency classification, and cost-efficient first-pass triage |
| Amazon Nova 2 Pro | Tool output summarization | Condensing MCP tool responses and generating readable summaries for analysts |
A major technical issue was context window management. The context window is the amount of text a model can consider at once. Duco exposed 300-400 tools per tenant through AgentCore Gateway, which fronts a separate MCP server for each underlying Duco service. Before work began, approximately 200,000 tokens of tool descriptions were loaded into context. This caused context overloading: response quality degraded as conversations lengthened, and the model could lose the thread of earlier reasoning. The team said a feeling that an agent is becoming less capable is not enough evidence on its own, and evaluations must be run before engineering decisions.
Duco addressed this with a skills architecture based on MCP resource types. Tool descriptions were moved out of static context and loaded only when needed.
- `skill://` resources load tool descriptions on demand when a skill is required.
- `system://` resources compose dynamic system prompt sections at runtime, keeping the base system prompt lean.
- `ui://` resources hold UI-specific material and are isolated from operational tooling.
Three new meta-tools were implemented in Duco's agent harness:
- `load_skill` fetches skill descriptions only when the agent determines a skill is needed.
- `find_tool` locates relevant tools without loading all schemas into context.
- `invoke_tool` executes the selected tool once it has been identified.
The source reports a 90% context reduction, with main context falling from 200,000 tokens to approximately 20,000 tokens by moving tool descriptions out of static context and limiting tool response size. Prompt caching was preserved because the static prompt prefix, made of the system prompt and the three meta-tools, stayed small and stable. Skills load into the conversation rather than the tool list, so loading a skill does not invalidate the cache. The team says agent response quality improved measurably, validated through a rigorous evaluation framework, and that the change required zero infrastructure changes: the same underlying models, tools, and tasks were used.
The source also describes MCP tool discovery at boot time. When a new service is deployed, its MCP server is registered as a target on the tenant's AgentCore Gateway as part of the deployment. The available material stops before giving the full discovery sequence that follows.
What the design pattern shows
For non-experts, the practical question is whether AI agents can be used in financial operations without mixing tenant data or overloading the model with too much information. Duco's architecture points to two patterns: hard isolation by tenant and on-demand loading of tools and skills. The result is a setup where each bank's agent can work only inside that bank's context, while the platform can still scale across 200+ clients and 20 billion monthly transactions.
Comments (0)