Back to the room

Legacy Modernization Assistant

A multi-agent system on Claude that reads legacy code such as COBOL and DYL-280, designs a modern replacement with human input, and generates a first cut in the stack of the customer's choosing. In use by the State of California.

Role
AI Architect
Type
Professional
Client
State of California

01

Overview

The Legacy Modernization Assistant (LMA) is a multi-agent system on Claude that takes a legacy system and modernizes it into a modern stack of the customer's choosing. It accepts any legacy stack as input and any modern stack as output.

I architected it end to end: the agent roles, the human review points, the security model that keeps source code inside the customer's cloud, and the validation that checks the new system against the old one. It is in use by the State of California.

02

The Challenge

Critical business systems built on legacy technologies such as COBOL and DYL-280 are difficult to maintain and extend, because the architects who built them have retired. The knowledge of how the system works lives mostly in the code itself.

Any tool that helps has to meet enterprise constraints: the legacy code cannot leave the customer's cloud, people have to stay in control of design decisions, and the new system has to be proven to behave like the old one.

03

The Solution

LEGACYCOBOL · DYL-280ANALYZEREAD · EXTRACTDESIGNHUMAN INPUTGENERATEANY STACKVALIDATETESTS · DIFFHUMAN IN THE LOOPSSOENTERPRISE AUTHENTICATIONCHANGE MANAGEMENTBUG FIXES · TEST GENERATION · DIFF ANALYSISENTERPRISE CLOUD · CLAUDE ON AWS BEDROCK OR AZURE AI FOUNDRYLEGACY MODERNIZATION ASSISTANT
Legacy Modernization Assistant: system architecture

I designed LMA as a pipeline of specialized agents with a person in the loop at the points where judgment matters.

Analysis agents

Agents read the legacy code, extract its structure and business logic, and analyze how the system behaves, turning code nobody fully understands anymore into insight a team can act on.

Design with human input

Agents design the new system and surface their findings for review. People make the design calls; the agents do the reading, the cross-referencing, and the first draft.

Generation in any stack

Once the design is agreed, LMA generates a first cut of the new system in the selected stack. The input and output stacks are both open: any legacy stack in, any modern stack out.

Change management

After the first cut, LMA supports day-to-day work on the system: bug fixes, test code generation, and diff analysis.

Governance and security

  • Human in the loop by design
  • Single sign-on enterprise authentication
  • Claude models run in AWS Bedrock or Azure AI Foundry, so the legacy code never leaves the enterprise cloud

Validation

Generated tests and output comparison check the new system against the original's behavior, so modernization is measured rather than assumed.

04

Results & Impact

  • In use by the State of California
  • Accepts any legacy stack as input and any modern stack as output
  • Legacy code never leaves the customer's enterprise cloud
  • Generated tests and output comparison validate the new system against the original