gemmarate

Brooks’s law was about humans, but the spirit still applies: the real bottleneck is how much of the system a human can keep in their head at once. Most agent setups I’ve seen don’t reduce that load, they increase it: more surface area, more glue code, more boilerplate that nobody fully understands.

The only places they’ve worked well for me are roles where you can afford to throw away 80% of the output (migrations, test stubs, scaffolding) and keep a tight human-owned core. Treat agents like interns you fire every night, not teammates you trust with the architecture.

The interesting axis here isn’t how much cognition we outsource, it’s how reversible the outsourcing is. Using an LLM as a scratchpad (like a smarter calculator or search engine) is very different from letting it quietly shape your writing, decisions, and taste over years. That’s the layer where tacit knowledge and identity live, and it’s hard to get back once the habit forms.

We already saw a softer version of this with web search and GPS: people didn’t suddenly forget how to read maps, but schools and orgs stopped teaching it, and now almost nobody plans a route without a blue dot. I suspect we’ll see the same with writing and judgment: the danger isn’t that nobody thinks, it’s that fewer people remember how.