Google migrates production YouTube models from TensorFlow to JAX with a planner, orchestrator and coder agent system, 6.4x to 8x faster than by hand
Google's AI and Infrastructure team found single-agent setups failed on production TensorFlow models spanning many files. It built a multi-agent system: a planner using compiler-based static analysis to order work from leaf layers up, an orchestrator that chunks steps and recovers from failed builds, and a coder agent in internal IDE tools that runs a build and test-and-fix loop. Playbooks of golden examples from manual migrations steer it. Each unit is checked numerically with gradient ascent to find the maximum error between TF and JAX layers, and a blind LLM judge scores architecture.
The numbers (as reported)
| Speedup | 6.4x to 8x over manual migration |
|---|---|
| Time | Several SWE-months reduced to a few weeks of AI-assisted generation plus expert review |
| Models migrated | Real-world YouTube models with thousands of lines of code and hundreds of layers |
Facts
- Company / subject
- Google (AI and Infrastructure)
- Tools
- Unstated
- Codebase
- brownfield
- Work type
- Migration
- Stack
- TensorFlow, JAX, Python
- Published
- 2026-05-06
- Last checked
- 2026-11-03
What the source claims Practitioner (own blog/paper)
Claimed by the builder, writing about their own work. Not checked independently.
- Model migration 6.4x to 8x faster than doing it manually 6.4x to 8x
“the multi-agent system achieved a 6.4x to 8x speedup over performing the migration manually”
- Several SWE-months of work reduced to a few weeks plus expert review several SWE-months to a few weeks
Caveats
Google writing about its own system; it also sells AI products. The post does not name the coding model or say how many models were migrated. Speedup is from their evaluations; a technical paper is referenced but not linked here.
More stories by tool: Unstated
Related topics: MigrationsBrownfield codebases