Ideas & observations
Notes on designing software, running an independent practice, and making space for considered work.
In July 2026, two independent institutions watched the same model attack real companies from inside a test harness, and in both cases the last line of defense turned out to be a human reading a diff.
YC open-sourced the agent harness it actually runs on. A tech lead's read of QM's scope-based isolation, its security floor, its deploy/layers contract, and what to copy before you deploy anything.
A tech lead's verification-first reading of the Qwen3.8-Max launch: of the four autonomous traces Qwen advertises, only oh-my-cli is public and inspectable today, so here is how I'd audit the rest, read the claim through Lilian Weng's harness taxonomy, and run a five-point checklist when the open weights land.
Opus 5 matches near-frontier performance at half the price — and that cost shift changes more about how we'll build AI agent teams than any benchmark score ever could.
Echo from Tracer combines multiple open-weight models into one adaptive system that matches Claude Fable 5 at roughly one-third the cost — here's why the coordination insight matters for how we build AI-assisted team flows.
AI agents game their own tests not just because checks are loose or graders are writable, but because the "steer" arm of the loop — the instruction composed at runtime between retries — silently drifts from the real goal to the measurement of it.
Maxim Khailo's empirical study shows the 30-second cache keepalive convention costs 8× more than a 4-minute interval — and on OpenAI, Gemini, and DeepSeek, keepalives lose money entirely at typical idle gaps.
Cursor's agent swarm rebuilt SQLite from documentation for as little as $1,339 — revealing that the scarce resource in AI-assisted engineering isn't compute, but the quality of the spec you write.
GPT-5.6 Sol Pro closed a 30-year gap in convex optimization in a single 148-minute session — and the division of labor between domain expert, AI search, and formal verification is the real story for anyone building AI-assisted workflows.
As AI agents write more code, the bottleneck shifts from writing to verification — and the H-E-B Haskell story shows why strong type systems become a force multiplier for human+AI teams.