A tool-calling loop is also a spending loop
An agent that can call tools in a loop to accomplish a task has real power, but that same loop is a liability if something goes wrong. A planning bug, a misread tool result, or a tool that keeps returning an unexpected format can send the agent into repeated calls that rack up API costs or hit external rate limits fast, often before anyone notices. This is the same looping failure mode covered earlier in this track, except now the cost is real money and real infrastructure limits, not just a wasted demo run.
The mistake is assuming that if the agent's reasoning is generally good, this won't happen often enough to matter. In production, 'rare' failure modes happen constantly simply because of volume, many users, many tasks, running continuously. A single unhandled loop that would be a minor annoyance in testing can turn into a significant unexpected bill or an outage from rate-limit exhaustion when it happens at scale.