Faye Zhang¶
Staff AI engineer and tech lead at pinterest, leading multimodal AI for search, traffic discovery, and shopping — driving platform growth globally.
At the Coding Agents Conference (March 2026)¶
Lightning talk Production Sub-agents for LLM Post Training (faye-zhang-subagents-posttraining-2026): compressed Pinterest's post-training cycle from 4–6 weeks to ~1 week by parallelising data generation, SQL injection, and training parameter sweeps via claude-code sub-agents. Honest about the failure modes introduced (see agentic-posttraining-failure-modes).
Grounded quotes¶
- "Before Claude code… it's a giant linear process… incredibly manual and takes 4 to 6 weeks."
- "Introducing our new workflow which just takes about a week using Claude code. We effectively break down the model generation of the training data using Claude code, SQL injection, and also training parameters. All of that task has been parallelized using sub agent structure."
- "The biggest difference is for sub agents, you essentially communicate with the sub agents only via your main agent."
- On the two bottlenecks: "Number one is context limitation… the context window quickly expand exponentially. Number two is the rigidness of the orchestration… the hot celebrity problem in system design."
- On fixes: "Reinforce the agent orchestration with agent SDK… open source agent orchestration tool… has the nice hook mechanism."