Rafal Wilinski is a founding engineer at Runlayer and the creator of Dynobase. His work spans serverless infrastructure, developer tools and the evaluation of AI agents operating through external services.[1]
Start with Agentic Evals Pyramid →
X · rafalwilinski Career
Wilinski’s timeline lists cloud engineering at X-Team, serverless engineering at Stedi, AI technical leadership at Vendr and technical leadership for Zapier Agents. He joined Runlayer as a founding engineer in September 2025. He also founded Dynobase, a graphical client for Amazon DynamoDB. His current project list includes ChatWithCloud, an AI command-line interface for AWS, and express-status-monitor for Node.js applications.[1]
Contributions
Wilinski describes creating disposable test identities with known data rather than substituting fake API responses. The calls and response formats stay real, while test email histories and contacts remain controlled. He places real-user A/B tests above offline evaluations, because a narrow rubric can miss changes in the experience. The resulting feedback becomes new lower-level regression cases.[2]
Ideas
In Agentic Evals Pyramid, Wilinski separates quick assertions about individual outputs from evaluations of an entire sequence of tool calls. An exact next-step test can penalize a model that reaches the right outcome through a different route. Trajectory evaluation instead examines the final artifact, necessary research, efficiency and recovery behavior, using ordinary assertions or a rubric where appropriate.[2]
Tools, writing & talks
Writing, research & talks
Start here
Agentic Evals Pyramidrwilinski.aiExplains how output assertions, full-workflow evaluations and user experiments complement one another.
Start with the evaluation essay if you are building an agent that calls APIs. Read the examples as a comparison of testing strategies; familiarity with software tests and tool calls helps.
Author[2]- From 36% to 100%: How Self-Improving Agents Write Their Own Skills — Rafal Wilinski, Runlayeryoutube.com
Runlayer's founding engineer shows self-improving agents that distill skills from successful and failed runs, distributed company-wide over MCP. One distilled skill raised a task from 36% to 100% success, building an organizational knowledge flywheel.
Catalog workOriginal source
Tools
- express-status-monitor
github.comA real-time monitoring solution for Node.js and Express applications.
Creator[1]