Zep AI
Day 28 after signup
The research behind Zep Agent Skills
How a stronger model taught cheaper ones. Zep Agent Skills is free until January 4.
Why it works
- Lead with the specific research question so readers quickly understand what the findings investigate.
- Explain the test setup before presenting the quantified results to make the evidence easier to follow.
- Pair the research link with a practical next step so interested readers can move from learning to trying the product.
How it is built
- Research header
- Brief product context
- Research method
- Benchmark results
Email copy
RESEARCH Hi [First name], Earlier this month we launched Zep Agent Skills. Today we published the research behind it. Our research tested whether a cheaper model can use what a stronger model learned. The teacher, Claude Opus 5.5, ran tasks from three benchmarks. Zep compiled skills from the passing runs, and cheaper models ran the same tasks with and without them. Skills raised the cheaper models' pass rates in all 26 setups across the three benchmarks, by 4.5 to 18.8 points. On EnterpriseOps-Gym, each model passed more tasks with skills at low reasoning effort than without them at maximum effort. Read the research The full paper is on arXiv. Get started. Zep Agent Skills is free on every Zep plan until January 4, 2027. Upgrade to the v4 SDK and send your agent's runs, or import them from LangSmith or Braintrust. Read the quickstart Daniel You are receiving this email because you opted in to receive email from the list Product Notifications from Zep AI Zep AI, 2261 Market Street STE 5686 San Francisco, CA 94114 Unsubscribe