6 October 2001 · 3 min read
Red-teaming your support bot: a field guide
Twenty things customers will try on your support agent in its first month, grouped by how they work and what a passing response looks like.
arenasecuritycustomer-service
Read Resources · Insights
Notes on practice, evidence and agents.
Short, specific pieces on how conversation practice and agent evaluation actually work. No listicles.
6 October 2001 · 2 min read
The multi-turn blind spot in agent evaluation
Your eval suite scores single responses against expected answers. Your customers run multi-turn conversations with goals and memory. The failures that hurt live in the gap.
arenaai-agentsevaluation
Read 6 October 2001 · 2 min read
The return-to-office conversation nobody practises
A top performer says 'I won't be doing three days'. Most managers argue the policy. The ones who keep the person ask a different question first.
managersscenariosmethod
Read 6 October 2001 · 2 min read
Why feedback must quote the transcript
Most AI feedback is confident, fluent and unverifiable. We built a judge that is not allowed to say anything it cannot point to — and a verifier that deletes what it can't find.
methodevidenceproduct
Read See it with your own scenarios.
Pilots start with your real policies, your real objections, your real customers — not generic demos.