A recent Hacker News post recounts a consultant who ran out of daily tokens and was told to “make do.” The typical reply—"you should’ve been able to pick up the work midway"—misses the reality of juggling dozens of (sub)agents, especially when models like Claude Fable 5 and Claude Mythos 5 hide their chain‑of‑thought. Engineers spend hours untangling workstreams, documenting manual steps, and hoping not to break internal consistency. When the limit resets they hand the LLM back its own output, only to watch it recreate hours of work in seconds, or they wait for something more useful.

Different roles demand different mixes of orchestration and coordination. While some jobs center on managing agents, others involve reading docs, meetings, or incident response. Companies still tie pay to hours, yet many also use objectives and deadlines. The expectation to fill the day with any work raises the question: can a team member stuck at 5h:100% 7d:100% study, learn, or explore colleagues' tasks? Many already do this noble activity, but cultural acceptance and incentives vary. The Uber case, where the year’s token budget was exhausted by April, shows that engineering still wrestles with constraint management.

If firms push for slower agent usage to respect contracted hours, they may under‑utilize AI and drag everyone down. Conversely, doubling tokens can encourage sloppiness, letting agents clean up later. The core tension is clear: engineers must squeeze the most useful LLM work into limited tokens, just as programmers once squeezed code into kilobytes. Are teams ready for this new budgeting challenge? The discussion continues.