What we learn building agents that resolve.
How they work under the hood, how they're tested, how to measure whether they really resolve and what changes for your team. Engineering, product and customer support, told with real cases.
Chatbot vs AI agent | The difference is whether the case gets closed
Both talk well and both can run on the same model. What separates them is what happens after the reply: whether something goes into your systems, makes the change and checks that it's done.
Kevin Manotas
October 12, 2026 · 5 min read
Evals for an agent that acts | Test what it did, not how it said it
How to test a support agent before publishing it: what a case contains, why we judge it with rules instead of another model, how actions are tested without touching a single customer, and why one failure stops the publish.
RAG for AI agents, and how we do it at Qopper
What RAG is, in plain words, how it works under the hood, and what it takes for an agent that also acts on your systems to never promise a number nobody wrote.
RAG doesn't know where your order is
The difference between what your business has written down and what's happening right now, why the second can't be indexed, and what it takes for an agent to look it up without making half of it up.
How an AI agent actually takes action in your systems
The model doesn't execute anything: it asks for it to be executed. What tool calling is, what the full loop looks like, and why reads and writes aren't treated the same when something fails halfway.
The AI layer for your customer support
Agents that resolve, not deflect: they act in your systems, verify and respond.