Agents should warn before launching if token budget is low

Problem

When launching multiple agents in parallel near the usage limit, all agents fail mid-task. The tokens they consumed are wasted — reads, greps, and analysis are lost with no partial results returned.

The user has no visibility into remaining budget, and Claude doesn't check before spawning agents. This is like starting a long drive without checking the fuel gauge.

What happened

User asked to launch 3 Plan agents in parallel. Each did 5-9 tool calls (file reads, searches) before hitting the usage ceiling. All three returned empty — zero usable output. The research had to be redone manually from the main context, doubling the total token cost.

Expected behaviour

1. Claude should know (or estimate) remaining token/usage budget before launching agents
2. If budget is tight, warn the user: "You have ~X tokens remaining. 3 parallel agents may not all finish. Run sequentially instead?"
3. If budget is clearly insufficient, refuse and suggest a cheaper approach (sequential agents, or doing the research inline)
4. Partial results: if an agent errors out mid-task, return whatever it gathered so far (file contents read, search results found) rather than discarding everything

Analogy

A car shows the fuel gauge and warns before a long journey. Claude should do the same with token budget — don't let the user burn resources on work that can't complete.