New Research Shows AI Agents Work Better When They Delegate

Imagine hiring a chef who reads every recipe in the cookbook before cooking anything. By the time they reach dessert, they've forgotten the sauce. That's roughly how today's AI assistants behave on long jobs. They hold everything in a limited working memory — called a "context window," basically a scratchpad that fills up — and the more you pile on, the sloppier their thinking gets.

A new research paper from a team of five scientists tests a fix. Instead of stuffing all the instructions for a task into one AI's scratchpad, you break the job into pieces and give each piece to a separate helper AI — called a "subagent" — with its own blank scratchpad. Each helper gets one clear job, does it, and reports back to the boss AI. Think of a project manager delegating to specialists rather than trying to do everything personally.

The helpers won. Across their tests, the delegation approach beat the old approach — but only under certain conditions. It worked well when each task had a clear definition of what goes in and what comes out, like a form with obvious blanks to fill. When the instructions were vague, the advantage disappeared.

There's a real tradeoff. Every time the boss AI checks in with a helper, that conversation costs extra computing power — which means money and time. So you're trading a smarter, more reliable process for a bigger bill. The researchers' broader point is simple but useful: it's not just what knowledge an AI has that matters. It's how you organize it and who you hand it to. For anyone waiting on AI that can grind through genuinely long, complicated work, this is a hint at the plumbing that needs to change first.