Yours has just had less time. A founder watched me walk through how I run the company on agents and messaged afterwards: "amazed and also quite daunted". The first word is fine, the second one is the problem. Setups that look impossible aren't a harder idea than yours – mine is a folder of markdown files an agent reads, conceptually identical to whatever you'd build on day one. What matters is the maturity of your context and your workflows. Ten months of catching contradictions and gaps separates the two, and there's no real shortcut. The structure is what you arrive at, not what you design. Blog post

The cleaner way to ask Claude Code a clarifying question. Normally a clarifying aside costs context – the question and the answer sit in the transcript until you /rewind them back out. /btw <question> never records the turn at all, so there's nothing to roll back. Tweet

The strongest agent setups hand off through documents, not memory. Came up in a group thread about Claude Design passing prototypes to Claude Code, but the pattern is broader. A reference doc holds the stable ‘contract’ – components, constraints, what good looks like. A spec describes the per-task ask. Two scoped agents read both: one writes the spec, the next builds against it.

On my radar

Google launched a Claude Code competitor at I/O. Antigravity CLI is the new command-line agent, alongside Gemini Spark (a 24/7 personal agent on 3.5 Flash) rolling to Ultra subscribers. Anthropic's monopoly on terminal agents just ended, and Gemini at 900M+ users is a different distribution game. Google blog

AI content is collapsing in Google search. Analysis of 220 domains that publicly scaled AI content: 39% lost more than half their peak organic traffic, 22% lost more than 75%. Same failure mode as bloated context – defaults degrade output you don't notice until someone else does. Substack

The case against monolithic orchestrators. An argument doing the rounds: the current wave of orchestration products creates do-everything agents that bloat context and quietly accumulate too much access. Decompose to tightly scoped agents handing off through documents instead. Same lesson as everywhere else – constraints make systems reliable. Essay