What the session covers
Traditional applications answer one request with one response. An agent is given a goal and then searches, plans, retries and calls tools until it decides it is done. That is useful, and it is also a cost risk, because a small misconfiguration can leave an agent calling the same tool repeatedly, pursuing a task with no valid end state, or asking the model to reason about the same problem without reaching a conclusion. All of this happens invisibly and quickly. The clip draws a parallel with advertising budgets: if you let the platform decide how much to spend, the bill can arrive at ten or twenty thousand dollars. A runaway agent behaves the same way, except it does not wait for the billing cycle.
What you’ll learn
- How agents differ from request-response apps in the way they consume tokens.
- The three failure patterns: repeated tool calls, tasks with no end state, and reasoning loops.
- Why the damage happens before anyone notices, and why a monthly invoice is too late.
- What a circuit breaker for an agent consists of: token caps per run, tool-call limits, loop detection, approval thresholds and automatic holds.
In their words
“Nobody now reviews the entire code of how the agent is written. A slight misconfiguration, and it can get stuck in a loop.”
“A runaway agent does not wait for the next monthly billing cycle.”
Key takeaways
- If your agents can act autonomously, they also need boundaries.
- Put the boundary where the calls happen, at the gateway, so it applies to every agent regardless of who built it.
More sessions

Webinar · 50 min
Taming runaway AI spending: how to control LLM and agent costs
A 50-minute session on why AI bills spike, how token pricing really works, the five levers that bring spend under control, and why those levers belong in a gateway rather than inside each app.

Conference talk
Emotional AI and voice forensics: DMWF Global 2025
Sheoli Biswas on the Emotional AI for Marketing panel at DMWF Global 2025 in London: what sentiment and voice analysis can read, where consumer trust breaks, and the controls that keep emotional data from becoming a liability.