I build agent systems
that work beyond the demo.

The problem

When the agentic Help Center handed a conversation to a human, the human started cold. The traveller had already explained the problem, the agent had already looked up the booking, and none of it survived the handoff. Travellers repeated themselves to a second person who could see less than the first.

The constraints

• Regulatory. Booking and traveller data could not be copied into a new store without a retention story, so "log everything and let the agent console read it" was not available.
• Existing tools. The agent's context was spread across several tools owned by different teams. No one team could change the shape of it unilaterally.
• Live traffic. Roughly 2.5M conversations a month, so migration had to be incremental. There was no cutover.

The alternatives

• Re-derive on handoff. Have the human console call the same tools again. Cheap to build and needed no new storage — but it doubled tool load at exactly the moment a conversation was already going badly, and it could not reconstruct why the agent had escalated.
• Pass a transcript. Hand the human the conversation and let them read it. Simple, and what most handoffs do — but it moves the work to the person, and a transcript does not say which booking was in scope.
• Persist a context record, keyed to the conversation. More to build, and a retention question to answer, but the escalation reason becomes explicit rather than inferred.

The choice, and the trade

We persisted a context record with an explicit escalation reason. The deciding argument was not effort: it was that the first two options make the handoff a guess. A human who can see the agent tried a refund and hit a policy limit is solving a different problem than one reading a transcript.

The cost we accepted: a new record with a retention policy, and the escalation reason becoming an interface that other teams now depend on — which means it cannot be changed casually.

Bringing other teams along

[TODO — the part only you can write. What actually moved the tool-owning teams: was it a review, a prototype that showed the console with context in it, one incident that made the cost concrete? Name the specific thing and the specific objection you had to answer.]

What I would do differently

[TODO — one honest thing. The escalation-reason taxonomy being harder to agree than the storage? Underestimating how many callers would depend on it? Whatever you actually got wrong.]