PithTrain introduces an agent-native MoE training system with ATE-Bench
Cohere · youtube · 2026-07-25
PithTrain proposes an agent-native MoE training stack
The talk introduces PithTrain, a compact Mixture-of-Experts training system designed around a new metric the authors call agent-task efficiency (ATE): the cost of using coding agents to understand, operate, and extend a training framework.
Key points
- The authors argue that optimizing for throughput alone misses a hidden cost when agents are involved.
- PithTrain is built around four agent-native design principles, including a compact codebase, a Python-native stack, no hidden indirection, and REPL-friendly agent skills.
- They also introduce ATE-Bench, a benchmark for real-world training-framework tasks.
- According to the evaluation, PithTrain matches production-framework throughput while improving agent-task efficiency.
- Reported gains include up to 62% fewer agent turns and 64% less active GPU time on ATE-Bench.
The speakers are Ruihang Lai and Hao Kang from Carnegie Mellon University, and the session was hosted by Cohere Labs.
More from coding & agent
- Claude Code article says removing 80% of the system prompt changed how newer Claude 5 models work — AccBalanced · 2026-07-25
- Codex workflow tip: draw the flow first, then ask it to implement it — alex_frantic · 2026-07-25
- Microsoft’s OpenForgeRL trains agents inside the same harnesses they use in production — dair_ai · 2026-07-25
- Paper argues production agents fail from context overload, not reasoning — omarsar0 · 2026-07-25
- Gemini CLI fixes Windows diff views by normalizing CRLF to LF in a2a-server — luisfelipe-alt · 2026-07-25
- VaynerX says a $20,000 internal app was prototyped in hours with Claude — tomcrawshaw01 · 2026-07-25