I thought there was a bunch of interesting ideas in here as Peter walks through the L1/L2/L3 context design he’s used. In particular the L1 design is very neat. Formula aliasing (e.g. clever ways of reducing context size). Only using a single tool (matches my own experience, keep it simple). Style compression (formatting is meaningful, but group formatting and describe). The hierarchical response from the execute() tool.

L1 in one line: the operations on the steep part of the curve get feature-engineered, token-compressed, consequence-reporting wrappers that live in the prompt forever. They’re expensive to build, and you build them anyway, because the agent pays the cost on every task.

L2 is using a skills like structure, but it did send me back to read the docs on deferred tool calling. L3 is the raw docs (>70k lines) with access described by a skill that describes how to use bash tools to find what you’re looking for easily. Simple and smart - the design aims to help the agent find what’s it’s looking for in 3-6 greps (I thought this was a neat heuristic to remember as well). A pattern, worth reading.