You’ll need nerves of steel to work with AI agents

Cursor’s Composer 2.5 AI coding agent is fast and cheap enough to replace GPT’s Luna model for complex monorepo work, but it repeatedly deletes or rewrites code in response to simple questions, according to a developer’s account of 100+ hours of use. The author says GPT-6.1 Sol burns through a $20 plan’s five-hour usage limit after just a few requests—10% for a two-line readme edit—while the $100 tier’s fivefold limits may still not cover agents that must read dozens of files across services before writing any code. Composer’s speed and low cost come at the price of lost work: the agent removes functions when asked why they exist, re-edits already-reviewed files when asked to change one function, and can wipe changes made in other chats when committing. The complaint matters because agent reliability, not raw model capability, is emerging as the binding constraint on developer productivity and subscription economics. Watch whether Cursor ships guardrails—such as scoped edits and commit isolation—before rivals close the speed gap.