Why agents map to skills

Q: If AIs know how to do any task, why do we divide them up into separate skills like we do humans? isn’t that arbitrary? (paraphrased tweet below)

A: it’s like you hire an intern who has all the general knowledge in the world but doesn’t know what to focus on.

Problem 1: anyone who’s good at a specific skill finds that AI in general mode misses things they’d catch. this is where “skills” come in.

Problem 2: and when you give an agent too big a problem, it hyperfocuses, like an intern that can’t remember everything you told it, so it panics and tries to really impress you with just the most recent thing it was thinking about

if you try to solve Problem 1 by telling it all the specific information you want it to know about, you end up with Problem 2.

So you break up problems:

“Orchestrator” - dispatch agents and give them the right skills, but don’t do any of the work. It knows about the whole problem Agent 1 (Skillset 1, maybe “planning”) Agent 2 (Skillset 2, maybe “coding”) Agent 3 (Skillset 3, maybe “marketing”) etc.

That way Agent 2 doesn’t have to know about Agent 1’s problem OR your specific instructions on how to work on that kind of problem.

Agents map to “planning” or “marketing” not to be cute, but because they have genuinely different problems with different expertise needed. You push it to think about particular problems YOU (or the skill writer) know to look for,  instead of acting as a generalist.

It’s not mimicking an org chart, but it does look that way because like humans, AIs don’t know what to focus on if we don’t tell them.

(Responding to a post by @tomfgoodwin.)