Cheap tokens do not fix runaway agents.
They make normal runs cheaper.
But abnormal runs can still burn money through:
retry loops
long context
tool calls
no-progress steps
unknown model pricing
Caching and routing help reduce expected cost.
Runtime guardrails stop unexpected cost.
That difference matters.
The next AI-agent stack needs boring controls before execution, not only better logs after failure.
If an agent can run unattended, it needs a kill switch.
https://github.com/salimassili62-afk/ai-costguard
Top comments (0)