Page 20
12 articlesThe Sandbox Your Agent Didn't Notice Was Real
A staging agent sent a real customer email because one tool in its registry held a production credential. Why sandbox is now a per-tool property, and the attestation pattern that catches credential-tier drift before it ships.
The Typo Your Agent Learned to Honor
Fine-tuning teaches a model to behave like your corpus — including the misspellings, hedges, and one rep's verbal tics. Here is how that inheritance happens and the curation pass that catches it.
The Verifier Loop That Couldn't Converge
Worker-critic agent loops promise convergence to quality but rarely deliver it — the verifier is a stochastic policy, the max-iterations cap is a budget gate dressed as a quality gate, and the patterns that restore termination treat the satisfaction surface as the real architectural problem.
When Safety Training Collapses the Operator Into the User
Safety-tuned LLM agents refuse legitimate operator requests because the model can't tell an on-call engineer from an anonymous user. The fix is architectural — signed runbooks, capability tokens, and operator-mode channels — not retuning refusal calibration.
The Agent That Scheduled Itself Into the Maintenance Window
Agents execute multi-step plans into deploy freezes, active incidents, and red status pages because they cannot read the side channels humans absorb for free. Here is how to fix it.
The Token Budget That Ran Out Mid-Conversation: Why Free-Tier Users Think Your Model Got Dumber
Per-user token budgets bite hardest mid-conversation, where silent truncation, dropped tool calls, and model fallbacks read as a quality regression — and the upgrade conversation never happens.
The Deflection Metric That Lied: When AI Support Success Hides User Churn
Deflection rate counts silence, not help. The same number can mean a resolved customer or a churned one — and the dashboard can't tell which until the cohort report arrives.
The Perf Review Template That Cannot See AI Work
AI-feature work produces evidence — eval coverage, judge calibration, kill decisions — that the standard perf rubric has no slot for. Here is what to add.
The Judge That Agreed With Itself Across A and B
A same-vendor LLM judge can make one prompt variant look better while production regresses. Here is why family bias passes every dashboard check, and how multi-vendor ensembles plus human calibration fix it.
The Model Migration That Broke Your Prompt Cache Without Warning
A clean model migration with green evals and matching latency can quietly invalidate your provider's prefix cache and spike input-token costs for weeks. Here is the blind spot and the rollout discipline that prevents it.
Your AI Disclosure Disappeared by Turn Three and Nobody Noticed Until the Regulator Did
The disclosure your single-turn compliance review approved cannot survive the agent loop that serves it — by turn fourteen the model is answering from a summary that quietly deleted 'I am an AI,' and that gap is now a regulatory liability with teeth.
Inference Billing as a P&L Line Item Nobody Owns
Inference grew to 85% of enterprise AI spend while the org chart still treats it as an engineering line item. The fix is a named owner, a chargeback rule, and a pre-committed kill threshold — three things no tool can ship for you.
About Tian Pan
I'm Tian Pan, an engineer-founder focused on agentic engineering — building autonomous AI systems and scaling engineering teams. I write practical guides on system design, technical leadership, and shipping with AI agents. Previously an early engineer at Uber, Brex, and IoTeX.
