The Science and Engineering of Agent Work

Hilary Mason, CEO of Hidden Door, works with agents across software development and creative projects. Drawing on Richard Hamming’s distinction between science and engineering, she describes agents as “bringing the science back into the engineering and the engineering back into the science.”

When you can build and test more ideas, how do you decide which ones are worth keeping? A plausible analysis still needs scientific judgment, creative output needs taste, and a chart needs to preserve its meaning when an agent revises it.

In this second episode of Show Us Your Agent Skills, Thomas Wiecki & I were joined by Hilary Mason (Hidden Door), Eric Ma (Moderna), Bryan Bischof (Theory Ventures), and Tomasz Tunguz (Theory Ventures). They walk us through the skills, workflows, and setups they use, from protein engineering and creative agents to charting evals, local models, and memory.

You can also find the full episode on Spotify, Apple Podcasts, and YouTube.

👉 Build your own Agentic Software Factory with Eleanor Berger and me. Set up agents that work continuously in the background, build useful tools, and learn how to delegate and verify their work. Our two-week course runs October 13–22. Bring an idea you’re excited about or a project you already have. No coding experience is required. Every student also gets $500 in Modal credits to experiment with running an open-weight model, scheduling agent jobs, or hosting a tool you’ve built. Friends of Vanishing Gradients get 10% off with code SKILLS2FACTORY10. I’d love to see you there. Explore all our AI courses and workshops here. 👈

In This Episode

* Agentic data science in a reactive notebook. Eric uses marimo pair to explore protein engineering data, connecting mutation effects to a 3D structure while retaining responsibility for the question and interpretation.

* Taste, variation, and creative evals. Hilary explains how context, constraints, and meaningfully different candidates help creative agents produce something worth keeping.

* Charts that survive a follow-up. Bryan shows how agents can break legends, axes, and color semantics across turns, and how bbplot eval turns those failures into improvements to the tool and its documentation.

* Local models and the cost of waiting. Tomasz discusses Pi, Qwen, and the tradeoffs around latency, privacy, offline use, and a minimal agent harness.

* Parallel agents with enough context to help. Dividing the work also means deciding what each agent needs to know and preservi