A tool timeout limits one request. An agent can still keep calling that tool—so it also needs limits on its overall execution and spending.
That’s one of the problems covered in my open-source coding-agent skills for Google ADK on Google Cloud. Other skills cover memory, deployment, authentication and evaluation, drawing on examples I’ve tested.
Repository and usage instructions
Which failure modes have you found hardest to handle in your ADK applications?
Let’s connect on LinkedIn: https://www.linkedin.com/in/ruslan-k-b6a48a1a6/.
Thanks for sharing this — very timely. I run a conversion-focused agency building Wix + HubSpot systems and experimenting with Symphony agents. The spending cap issue is huge for client trust. Have you tested patterns for evaluation that work in production, not just benchmarks? Would love to see your repo.
Hope it will help you. There is certain skills that help to control spending. I’d recommend starting with /adk-operational-guardrails and ask to create daily allowance per user per project to prevent burning too much tokens. I did indeed used what in those skills in production.