AI With a Manager Writes Better Code at a Fraction of the Cost

A new paper tests whether two AIs are better than one at coding. The researchers set up “manager-worker” teams: one AI acts like a project manager, breaking a coding problem into smaller chunks and reviewing the workers’ output. All the AIs share a “ledger” — basically a running set of notes that keeps everyone on the same page. This works with off-the-shelf models and requires zero extra training. On 100 hard coding challenges, the team approach helped some models a lot, while others performed the same or even worse.

Why should you care? Because the trick can make cheaper AI models nearly as good as very expensive ones. For example, one mid-tier model with a manager almost matched a flagship model’s score on the coding test — at one-fifth of the price per run. Another model is open-weight, meaning any company can host it on its own servers; that setup cost just $51.75 per 100 coding problems. For software teams watching budgets, this is a big deal: you might not need to buy a giant, costly AI if you can coordinate a smaller one.

The paper also explains why the manager helps. Two mechanisms stood out: context management (workers share notes so they don’t forget what’s already been done, which reduces errors) and problem decomposition (breaking a large problem into smaller pieces makes it easier for the AI to handle). But the benefit is not free — the manager approach roughly triples the token bill (the units of computation the AI consumes). So it’s a trade-off: fewer mistakes, but more computing power spent behind the scenes.

The big-picture lesson? Sometimes coordinating multiple AIs beats buying one smarter AI. As companies pour money into ever-larger models, this research suggests that better teamwork — not just bigger brains — could be the next wave of AI improvements. For regular users, that might mean faster progress and lower costs from AI coding assistants in the near future.