# Stratamend — Legacy code modernized by supervised AI agents > Canonical: https://stratamend.com/ > Last updated: 2026-10-08 Legacy code, modernized by supervised AI agents. Founded in Singapore in January 2026, Stratamend is building an agentic platform that maps, tests and migrates COBOL and aging Java into modern services. The work is done by Claude Code agents, verified against the legacy system’s real behaviour and approved by your engineers. - Built on Claude Code - Runs in your cloud - Human-approved changes ## Critical systems run on code few people can still change Banks, insurers and public agencies depend on systems written decades ago. Full manual rewrites are slow and risky, while keeping the old code running gets harder every year. - **220B**: Lines of COBOL estimated to still be in use worldwide (Reuters, 2017). - **Knowledge is retiring**: The engineers who understand these systems are leaving the workforce, and the business rules they carry are rarely written down. - **Rewrites stall**: Big-bang rewrites take years, freeze new features and often fail to reproduce what the old system really did. ## Modernization is an agent-shaped problem Migrating legacy code is a long chain of reading, changing, testing and fixing. That loop is exactly what autonomous agents are good at, and what autocomplete-style tools cannot do on their own. - **Acts, not just suggests**: An agent reads the repository, edits files and runs the build itself. A copilot waits for a developer to accept each line. - **Verifies its own work**: Every change is checked against tests captured from the legacy system. When tests fail, the agent reads the failure and tries again. - **Stays on long tasks**: Migrating a module means hundreds of steps across many files. Agents keep the plan and context for the whole job, not one prompt. - **Knows when to stop**: Agents work inside strict boundaries and hand ambiguous business rules to a human instead of guessing. ## How it works: Map. Pin. Migrate. Ship. Agents do the repetitive work. Tests prove the result. People make the decisions. 1. **Map**: Agents index the codebase and produce an Atlas: dependencies, dead code, embedded business rules and a risk-ranked migration order. 2. **Pin behaviour**: Before anything changes, we record what the current system actually outputs and turn it into golden-master tests. 3. **Migrate**: Claude Code agents rewrite one module at a time to a modern stack and iterate until every golden test passes. 4. **Review & ship**: Each change arrives as a small pull request with the agent’s reasoning. Your engineers decide what merges. ## Autonomous where it helps, controlled where it matters We are designing for regulated industries from the start, with Singapore’s MAS Technology Risk Management Guidelines and PDPA in mind. - Designed to run inside your own cloud account, calling Claude via Amazon Bedrock or Google Cloud Vertex AI. - Least-privilege agents: per-module tool allow-lists and hooks that block out-of-scope edits. - No agent can merge code. Every change needs passing tests and human approval. - Agent sessions, diffs and test results are logged for audit. ## Roadmap: What we are building, in order Each phase is delivered with design partners on real, sandboxed codebases before the next one starts. ### Phase 1: Atlas & test harness - Codebase mapping for COBOL and Java - Golden-master test generation from recorded runs - Migration plan export for engineering leads ### Phase 2: Supervised migration - Single-agent module migration with Claude Code - Pull requests with rationale and test evidence - Policy hooks for scope, secrets and dependencies ### Phase 3: Parallel fleet - Many agents in isolated git worktrees - Session audit log export for risk teams - Deployment inside customer cloud accounts ## Questions we hear often ### Is Stratamend available today? We are an early-stage startup, founded in January 2026, working with a small group of design partners. Partners get hands-on access as each phase of the roadmap ships and help shape what we build next. ### Why build on Claude Code? Modernization needs an agent that can read a large repository, run builds and tests, fix its own failures and work for long stretches without a person driving it. Claude Code is built for exactly that, runs from the terminal and in headless mode, and supports hooks that let us enforce hard boundaries on every action. ### Does our source code leave our environment? It does not have to. Our architecture is designed to run inside your own cloud account and call Claude through Amazon Bedrock or Google Cloud Vertex AI in your chosen region. ### What happens when an agent gets something wrong? It cannot merge anything. A change has to pass the golden-master tests, static checks and a human review. Agent sessions are logged so every change can be traced back to how it was produced. ### Which languages come first? COBOL and older Java (6–8) are the first sources we focus on, migrating to Java 21. Other languages follow based on design-partner demand. ## Help shape agentic modernization while it is being built. We are looking for a small number of engineering teams with a legacy system they want to modernize. You get: - Free early access to each phase as it ships - A codebase Atlas of one repository - Direct access to the founding team We ask for: - A sandboxed repository we can work on - A short feedback call every two weeks Apply at https://stratamend.com/#design-partners or email admin@stratamend.com. ## Engineering notes - [What we look for in a design partner](https://stratamend.com/what-we-look-for-in-a-design-partner.md) (2026-10-01) - [Keeping source code inside your cloud account](https://stratamend.com/keeping-source-code-inside-your-cloud-account.md) (2026-09-10) - [Java 6 to Java 21: the migration nobody budgets for](https://stratamend.com/java-6-to-java-21-the-migration-nobody-budgets-for.md) (2026-08-13) - [Writing PreToolUse hooks that keep an agent in scope](https://stratamend.com/writing-pretooluse-hooks-that-keep-an-agent-in-scope.md) (2026-07-09) - [How we plan to run many Claude Code agents without collisions](https://stratamend.com/how-we-plan-to-run-many-claude-code-agents-without-collisions.md) (2026-06-11) - [COBOL numeric semantics that break naive translations](https://stratamend.com/cobol-numeric-semantics-that-break-naive-translations.md) (2026-05-14) - [AI-written code in regulated industries: what an audit trail needs](https://stratamend.com/ai-written-code-in-regulated-industries-what-an-audit-trail-needs.md) (2026-04-21) - [Copilots versus agents for legacy migration](https://stratamend.com/copilots-versus-agents-for-legacy-migration.md) (2026-03-26) - [Mapping a legacy codebase: what goes into an Atlas](https://stratamend.com/mapping-a-legacy-codebase-what-goes-into-an-atlas.md) (2026-03-05) - [Why golden-master tests come before the first agent edit](https://stratamend.com/why-golden-master-tests-come-before-the-first-agent-edit.md) (2026-02-12) - [Why we started Stratamend](https://stratamend.com/why-we-started-stratamend.md) (2026-01-22)