Code-first analysis: make the agent write a script that returns only results instead of reading files into context

GitHub·medium signal

Rather than pulling large files or logs into the window for the model to read, have the agent write and execute a small script whose stdout is the only thing that returns — one execution replaces several read/grep tool calls and the raw bytes never cost context. This inverts the default 'read then reason' loop and compounds with tool-output sandboxing; it's the single highest-leverage habit for keeping long agent sessions cheap.