- The reason to understand code agents write is not just verification, it is participation. If you do not understand the system, you lose the ability to have the next idea. You can carry that cognitive debt for a while, but it eventually bites. geoffreylitt.com/2026/07/02/understanding-is-the-new…
- Analysis is breaking a problem into clean local pieces. Synthesis is making the pieces fit together again, which is why it is harder. For example, in engineering incident response is synthesis work: you only understand what is going wrong if you understand how the system normally hangs together. surfingcomplexity.blog/2026/07/03/synthesis-is-harder-than-analysis
- Bun’s Rust rewrite (the JavaScript runtime Claude Code runs on) is a good example of what AI can do when the work is shaped into tight loops. About 50 Claude Code workflows ran for 11 days as coding and review agents, but the important bit was the structure: mechanical porting, split implementer and reviewer contexts, compiler feedback, tests, fuzzing, and human side-by-side review. bun.com/blog/bun-in-rust
- Sierra’s agent lesson is that the useful shape is not one bot per function, but one agent across the job to be done. Collapse it into a single surface, make it proactive, give it the business context, keep the system of record as the backend, and measure whether outcomes changed sierra.ai/blog/ai-pilling-our-company-lessons-learned
- We should all learn something new. It makes life richer, the skills compound quietly, and you become more interesting to talk to. The useful warning is that early practice feels bad because your brain is gathering data for sleep to process later. marginalia.nu/log/a_135_learn
What I’ve been reading (W28, 2026)
Understanding, analysis vs synthesis, AI coding loops, company agents, and learning new things