Google uses LLMs for internal code migrations and reports about 50% less migration time

Google engineers published papers on using LLMs to run large internal code migrations, such as JUnit3 to JUnit4 and int32 to int64 IDs. Most of the landed code changes were generated by the model.

Organization: Google
Industry: Technology
When: Jan 2025
Evidence: Research paper
Agents and tools: Gemini (internal)

Reported outcome

Across 39 migrations, about 74% of code changes were LLM-generated; developers estimated a 50% cut in total migration time.

39
Migrations studied
~74%
LLM-generated changes
~50% (dev estimate)
Time saved

Caveats

Described in a paper with a stated method. Peer review status varies.

Time savings are developer estimates, not a controlled experiment.
Results differ a lot between migrations.

What the agent does

A pipeline finds the code that needs changing, asks the model to make the change, then runs builds and tests. Engineers review and land the changes. For one JUnit3-to-JUnit4 migration, reports say 5,359 files moved in three months, and about 87% of the AI code was committed unchanged.

What to copy

Use classic tools (search, static analysis) to find the targets. Use the LLM to make the edit.
Keep the change small per file, so review is fast.
Measure the share of model edits that land unchanged. It is a good quality signal.

Starter prompt

Written by us.

Change every use of {old_type} for {field_name} to {new_type} in this file.

Also update:
- function signatures that pass this field
- tests that build this field
Do not change anything else. If a change would affect an external API,
leave a TODO comment instead.

File:
{file_contents}