Direct execution beats planning at both token budgets tested

Direct execution beat planning at every token budget tested, per arXiv 2609.20449, a matched-budget comparison of agent workflows. The paper scored coding agents against externally verified software engineering tasks, holding the token budget equal so neither approach got extra room to work.

At a 12,000-token ceiling, direct execution scored 59.6% and information-constrained planning scored 36.2%, a 23-point gap. That's the cost of splitting an agent into a planner and an executor before the budget can support the split.

Double the ceiling to 24,000 tokens and direct execution didn't move, still 59.6%. Planning climbed to 51.2%, narrowing the gap to 8.4 points. The paper's bootstrap estimate puts that 15-point narrowing at a 95% interval of 4.2 to 25.8, so the effect holds but isn't tight.

One detail changes the math further. Letting the planner read the task issue directly is worth about 16 points at the 12,000-token budget, the paper found.

I split agents into planner and executor by reflex on most of my personal projects. This data says that split is a tax at low budgets. It starts paying off once there's enough context for the planner to see the real task, and enough tokens left to act on it.