Human-Agent Collaboration in Code Migration Human-Agent Collaboration in Code Migration | Zhiwei Li

Human-Agent Collaboration in Code Migration | Generated by AI

Home 2026.04

Looking back at our session — I delegated exactly once at the start via sessions_spawn (the sub-agent did the bulk CI3→CI4 conversion: 25 controllers, 32 models). After that, everything else I did directly with my built-in tools:

  • File editing (edit, write) — fixing broken define() calls, syntax errors, namespace issues
  • Reading (read, exec) — diagnosing errors, checking file contents
  • Shell (exec) — running curl tests, docker cp, grep, sed for bulk fixes
  • Process debugging — tracing stack traces, fixing CI4 DB API calls (result() → getResult())

The pattern that emerged: the coding agent handled the big upfront migration, and I handled the targeted debugging that followed. Once I saw an error, I could fix it in 1-3 tool calls without spinning up another agent.

So to answer directly: I delegate for large, unstructured tasks (building, mass refactoring, exploring an unfamiliar codebase). For known, targeted fixes (like fixing a specific error or updating a specific file), my direct tools are faster and more precise.

Back Donate