Attention budget
Also known as: human attention cost, attention cost per agent action
When a coding agent completes a task but requires five approvals, three clarifications, and a manual merge, the human time spent managing the agent may exceed the time saved. Attention budget is the idea that this cost should be designed for explicitly. The goal isn't maximum autonomy for its own sake; it's the lowest possible interruption rate that still keeps humans in control of outcomes that matter.
The term surfaced prominently in DeepSeek's October 2026 design document for dsh-superagent, which argued that agent teams were optimizing the wrong thing. Benchmark pass rates measure correctness but say nothing about how much engineering attention the agent consumed along the way. An agent that finishes tasks without pinging the team is genuinely more useful than one that aces evals but generates noise.
Practically, designing for a tight attention budget means being selective about when to ask for confirmation, batching questions rather than interrupting once per ambiguity, and defaulting to an observe mode that records what the agent would have done without actually doing it. It's closely related to progressive delegation: you earn the right to take more autonomous action as trust builds, rather than starting at maximum autonomy and backing off after incidents.