How does Buzz actually handle context-window limits and API token costs for workspace agents?

[d-damba]

Running AI agents in a channel seems like it could get expensive fast. When an agent is permanently embedded—reading chat history, reviewing code changes, and running multi-step tasks, it seems token usage could blow up.

I'm curious how the architecture keeps this overhead in check:

1. Context Filtering: Is there an automated way to prevent agents from resending the entire channel history on every single turn?
2. "Branch as Room" Scope: Does this setup strictly isolate the agent's context to the immediate branch/task, or can it still pull the global repo state?
3. Cost Gating: Are there any mechanics to limit token usage per agent per month for instance?

Any insight into how Buzz tackles these cost constraints would be greatly appreciated!

Thanks