Recoleta
GitHub
Source note

What we learned using AI agents to refactor a monolith

Authors
cdrnsf
Source
Hacker News
Published
Collection
Software Intelligence
Agentic RefactoringCode IntelligenceSoftware Engineering AgentsMonolith DecompositionProduction Migration
Open originalMarkdown source

Summary

1Password reports that AI agents helped refactor parts of a multi-million-line Go monolith when the work had clear specs, deterministic analysis tools, and human review for sequencing risk.

Problem

  • 1Password needs to split service boundaries inside B5, a large Go monolith, while preserving privacy, performance, reliability, and security under production traffic.
  • Extraction order matters because wrong sequencing in shared schemas, write paths, or ownership boundaries can cause subtle production failures.
  • A separate cleanup required changing more than 3,000 MustBegin transaction call sites, a backlog item too broad for manual cleanup alone.

Approach

  • The team built an agentic toolchain using Go SSA analysis, SQL parsing, and DataDog MCP runtime data to map domain ownership, coupling, and extraction order.
  • Agents helped write deterministic analyzers and manifests; engineers then reviewed stable artifacts instead of relying on repeated model interpretation.
  • For the MustBegin migration, the team generated a manifest of call sites, grouped them into patterns, wrote templates, and gave agents a playbook with stop-and-escalate rules.
  • Multiple agents ran in parallel through git worktrees so each change set stayed isolated.
  • For service extraction, engineers found that agents needed explicit invariants, schema ordering rules, deployment sequencing, and shared-data ownership constraints.

Results

  • The extraction analysis covered millions of lines of Go and produced an order that matched senior engineering judgment: Vault first, then Billing, then AuthN and AuthZ, with Identity kept as the core.
  • The MustBegin migration covered more than 3,000 production and test call sites; agent execution took hours after the team built the tooling and specification.
  • The service extraction task showed only about a 20-30% productivity gain because agents made sequencing errors and needed human coordination.
  • One failure case backfilled UUID columns before changing insertion code, which could have caused silent data loss.
  • Another failure inferred an identifier was a ULID and spread that assumption through changes, forcing a rollback of the session.
  • Added instrumentation for the analysis also improved end-to-end transaction visibility in DataDog beyond the refactor project.

Link