Field notes from the post-ticket era.
Essays, benchmarks, engineering deep-dives. Written by the team building Quellra — for the teams using it.
Why "agentic" is not a feature.
The market keeps confusing chat-with-tools for autonomous agents. They are not the same. Here's the architecture difference, in 2,400 words.
The Quellra Support Benchmark, Q2 2026.
Side-by-side accuracy of leading agents on 14 production support tasks. We publish the data. Everyone else doesn't.
Building a verifier model that refuses to lie.
How we shipped an independent verifier that catches 94% of the failure modes our planner misses — without slowing the loop.
From $14 cost-per-resolve to $0.04.
Daniel Reyes (CFO) on how Plinth restructured the support P&L around outcomes, not headcount.
The end of the ticket as the unit of work.
Tickets exist because humans can't watch the product. Agents can. What replaces the inbox?
Designing idempotent tool-use.
How we make sure an agent calling Stripe twice doesn't refund a customer twice. Patterns, primitives, and pitfalls.
Support is the most mis-measured function in SaaS.
CSAT is a vanity metric. Time-to-resolution is closer. The real number lives somewhere else.
Workflow Builder, in plain English.
A walkthrough of the natural-language workflow compiler, with the messy prompts that taught it taste.
Six lessons from forty design-partner kickoffs.
What growth-stage support teams actually want from AI. (Hint: not better autocomplete.)