Agents and budgets

An agent is a function that can carry its own budget and call tools or other agents. A budget is lexical, shared with child work, and returns Exhausted when it runs out.


agent classify(ticket: str) -> str:
    budget: max_tokens = 2000, max_steps = 4
    result: str = analyze(ticket, "Classify this support ticket")
    match result:
        case Ok(answer):
            return answer
        case Exhausted(meter):
            return f"budget exhausted: {meter}"
        case Uncertain(reason):
            return f"needs review: {reason}"
        case Failed(why):
            return f"provider failed: {why}"

Set a shared budget around a workflow:


def main():
    with budget(max_tokens = 10_000, max_calls = 20):
        results = parallel for ticket in tickets:
            return classify(ticket)

Each branch has an isolated heap, and results return in input order. Add max_time or max_seconds when a workflow needs a deadline. Inspect usage with tokens_spent(), tokens_remaining(), and calls_spent().