Sid Sethi

addendum: "but how do you do without websearch?" first, i'd like you to proxy your harness of choice and see what garbage their respective web search/fetch tool actually pushes into your agent's context. second, 99% of the time, the only reason you let your agent use web search is to fetch docs. these may be outdated or incomplete. i stead, i let my agent clone the repo of the dependency, analyse its code (or docs if available in the repo in their raw, non-html-rendered form) and optionally let it spit out an .md file that summarizes what it needs to know for future use. less token waste, more precise, reusable and updatable. "but my subagents" knocm yourself out. for me, subagents are only useful to: - gather context - try out things and throw them away in both cases i use pi's /tree instead. i have full observability, can take any learnings with me to a previous poibt in the session via summarization or writing a markdown file on disk. bonus: all the branches i took stay in the same session file, should i ever need to revisit them (i usually don't, sessions are generally ephemereal for me)