Agents
Containment vs. escalation: designing agent handoffs
Listen — audio summary in Nehal Vyas's voice
Transcript
Hi everyone, I recently shared my thoughts on designing effective AI agent handoffs, specifically how we balance containment and escalation to build more trustworthy AI. My biggest takeaway is that while containment – an agent fully resolving a request without human help – is often seen as the headline metric, chasing it blindly is how you end up trapping users in frustrating loops. A conversation that's "contained" but leaves a user unhappy is actually worse than a timely escalation. It burns trust twice. That’s why I always recommend measuring containment alongside downstream satisfaction and repeat-contact rates, so we're never optimizing a number that hides a bad experience. A second key point is to escalate on uncertainty, not just on failure. The best agents are designed to hand off proactively when their confidence drops, user sentiment turns, or the request involves a high-stakes action. These aren't emergent behaviors; they're explicit guardrails we design. And finally, make that handoff truly seamless. The human agent should never have to ask the user to repeat themselves, because all context – transcript, intent, and what the agent already tried – travels with them. That continuity turns a potential hiccup into a helpful feature. Ultimately, high containment is the outcome of trustworthy escalation, not a substitute. Build those exits well, and the agent earns the right to handle more. For all the details, you can find the full article on hinehal.com.
Containment - the share of requests an AI agent resolves without a human - is the headline metric for conversational AI. But chasing it blindly is how you end up with an agent that traps frustrated users in loops. The teams that win treat containment and escalation as two sides of one design problem.
The cost of the wrong containment
A contained conversation that leaves the user unhappy is worse than an escalation. It burns trust twice: once in the failed resolution, and again when the user has to start over with a human. I measure containment alongside downstream satisfaction and repeat-contact rate, so we're never optimizing a number that hides a bad experience.
Escalate on uncertainty, not just on failure
The best agents escalate before they fail. When confidence drops, when sentiment turns, or when the request touches a high-stakes action, handing off early preserves the relationship. Designing these triggers is product work: it's a set of explicit guardrails, not an emergent model behavior you hope for.
Make the handoff seamless
A clean escalation carries context with it. The human should never ask the user to repeat what they already told the agent. Transcript, intent, and attempted resolution travel together. That continuity is what turns a handoff from a failure into a feature.
High containment is the outcome of trustworthy escalation, not a substitute for it. Build the exits well, and the agent earns the right to handle more on its own.