Putting a dollar cap on one ADK session, using LiteLlm

[domondi1]

max_llm_calls limits how many calls an ADK session makes, but not what they cost, and sub-agents or a failing tool can still run up a bill on one request. Here's one way to give each session a hard dollar budget.

ADK's LiteLlm model takes api_base and extra_headers, so the session's id and budget can go to a small gateway that runs in the same process:

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

from google.adk.agents import LlmAgent
from google.adk.models.lite_llm import LiteLlm

def agent_for_session(session_id: str, budget_usd: str) -> LlmAgent:
    model = LiteLlm(model="openai/gpt-4o-mini", api_base=base_url, api_key="unused", max_tokens=200,
                    extra_headers={"X-Inferrail-Attribute-Work-Id": session_id,
                                   "X-Inferrail-Budget-Usd": budget_usd})
    return LlmAgent(name="support", model=model, instruction="...", tools=[...])

Each call reserves its worst-case cost before it's sent, and a call that doesn't fit is refused with a 402 before it reaches the provider. In ADK 2.x that arrives as an event with error_code set rather than an exception, so stop reading events at that one (the runner raises if you keep iterating). Sub-agents share the budget if they use the same session's model. inferrail work <session id> shows what it cost.

Caveats: only Chat Completions models through LiteLlm, not native Gemini; set max_tokens. Tested with google-adk 2.10.0 and 2.11.0, litellm 1.103.2.

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.