Break a large monolith into smaller modules
Agents take one slice of a big refactor at a time, following a playbook, while engineers review each slice before it merges.
The use case
Splitting a monolith is a long, repetitive project that rarely gets prioritized. Nubank's story with Devin describes moving a large ETL codebase into smaller sub-modules by breaking the work into many similar tasks, teaching the agent the pattern with examples, and running many sessions in parallel. Engineers reviewed and merged each piece and fed corrections back so later runs improved. This fits work where every unit follows the same recipe. Engineers still own the playbook, review every change, and keep checks in place to catch dependency cycles or broken imports.
Example workflow
- Write a migration playbook with worked examples
- Split the codebase into independent units of work
- Run agent sessions on units in parallel
- Review each change and note corrections
- Update the playbook and continue with the next batch
Tools
- Devin
- GitHub
- CI
Risks and guardrails
- One wrong pattern repeated across hundreds of changes
- Review fatigue on large volumes
- Cost can climb with many parallel sessions
Sources
Written by Agents Lookup. Credits and official pages for the tools this use case names:
- Cognition: Nubank customer story: devin.ai/customers/nubank
- Devin (official page): devin.ai (checked 2026-11-19)
- GitHub MCP server (official page): github.com/github/github-mcp-server (checked 2026-11-09)