Putting a dollar cap on one AgentScope run

[domondi1]

Multi-agent AgentScope runs can make a lot of model calls for one task, and there's no per-run dollar limit built in. Here's one way to add one without changing your agents.

OpenAIChatModel passes client_kwargs to the OpenAI client, so each run's model can carry its id and budget as headers to a small gateway in the same process:

import inferrail
base_url = inferrail.start()  # gateway on a background thread, no second process

from agentscope.credential import OpenAICredential
from agentscope.model import OpenAIChatModel

def model_for_run(run_id: str, budget_usd: str) -> OpenAIChatModel:
    return OpenAIChatModel(
        credential=OpenAICredential(api_key="unused", base_url=base_url),
        model="gpt-4o-mini",
        parameters=OpenAIChatModel.Parameters(max_tokens=200),
        client_kwargs={"default_headers": {"X-Inferrail-Attribute-Work-Id": run_id,
                                           "X-Inferrail-Budget-Usd": budget_usd}})

Give every agent in the run the same run's model and they share one budget. Each call reserves its worst-case cost before it's sent, so agents working at the same time can't overspend together. A call that doesn't fit gets a 402 before the provider (openai.APIStatusError, status_code == 402). inferrail work <run id> shows what the run cost.

Caveats: Chat Completions only, set max_tokens, only calls through that model count. Tested with agentscope 2.0.9 and inferrail 0.4.10 or later (earlier versions conflicted with AgentScope's mcp pin).

Full setup: guide. I maintain Inferrail (open source, Apache-2.0), so take the recommendation with that in mind. Curious how others bound the cost of a single run here.