Auto-adjust reasoning effort based on task difficulty

[miglaginf-tech]

Today the effort level (low/high/max) is chosen manually by the user, with no guidance from the agent. In my real usage, the cost is almost entirely concentrated in reasoning tokens, and a large portion is "over-spent" on tasks that don't need it.

My proposal: before starting, the agent should recommend the level based on the nature of the task (e.g., technical OCR + complex document with figures and formulas → high/max; summary, formatting, or ordinary conversation → low), and afterwards confirm whether the level used was appropriate (or whether it was too low / overkill). Ideally, it could auto-adjust in subsequent steps within the same session, since the technical mechanism for changing effort between steps already exists.

Reference case: simple tasks measured ~60% of output tokens as reasoning, when low would have been enough. Well calibrated, this would noticeably cut cost without affecting quality on hard tasks.