What an economics team delegates to AI coding agents, what they never delegate, and why identification stays with the scientist.
Sometimes the most important thing is learning how to iterate fast, in order to fail faster and more often with little penalty.
Why evals matter more than the skill itself, illustrated with a small personal workflow.
A mental model and decision tree for building agent skills incrementally: start with intent, add deterministic tools, then use tests and AI evals to reduce drift and risk.
A pain-driven approach to AI evals: measure first, constrain risk, iterate fast, optimize last.
Turn pain points into repeatable questions and build a lightweight taxonomy of your world.
How chaining LLM calls compounds errors, and why replacing probabilistic steps with deterministic ones improves reliability.
text/html
pip install
files.pythonhosted.org
