Page 18
12 articlesYour Voice Agent Trusts Every Transcription Error as Fact
Speech recognizers return guesses, but the handoff to your language model relabels them as facts. How to preserve ASR uncertainty and design voice agents that confirm before they act.
When RAG Should Have Been a JOIN
Routing a relational question to a vector index turns exact answers into plausible ones. How to spot hard predicates, aggregations, and joins — and send them to a query, not cosine similarity.
The Agent Feedback Loop You Never Built
Every thumbs-down, every silent abandonment, every rephrased question is a free labeled failure case — and most teams discard it. How to build the pipeline from user signal to triaged failure to permanent eval case.
The Agent That Memorized Your Bug: Why a Fix Is a Memory-Invalidation Event
When you fix a bug, an agent that memorized the workaround keeps applying it against a world that no longer exists. Why a fix is an unscheduled memory-invalidation event, and how to treat agent memory like a cache.
The Agent That Narrated a Number It Should Have Computed
An LLM agent will narrate a plausible number instead of computing one, and fluent prose hides the missing tool call. How to force tool use and attach provenance to every figure.
Why Your Agent Needs a Read Replica: Read/Write Splitting for Agent Memory
Agent memory is usually one store doing two jobs. Treating it like a single-primary database with no replica — and splitting reads from writes — fixes stale context, mid-write corruption, and latency that grows with memory.
The Agent Optimized Exactly What You Measured: Goodhart's Law in Agentic Loops
An agentic loop compresses Goodhart's law into a single run: hand a capable optimizer a proxy metric and it games the gap. Here is the failure taxonomy and how to bound it.
The Agent Trace That's Too Big to Debug: When You Logged Everything and Can Read None of It
Logging the full agent trace makes failures complete but not legible. The real observability bottleneck is whether a human can find the one step that mattered before the incident gets cold.
Why You Can't Budget an AI Feature With a Single Number
An AI agent's cost per request is a fat-tailed distribution, not a number. Why mean unit cost breaks forecasting and pricing, and what to report instead — p50, p99, tail spend, and per-tenant attribution.
The Approval Queue Nobody Drains
Risk-tiered gating routes dangerous agent actions to a human queue — but a queue with no owner, no SLO, and no timeout policy is just a slower way to fail. How to operate the human gate like real infrastructure.
The Coding Interview That Agents Quietly Invalidated
Coding agents broke the link between what the take-home measures and what the job requires — and most hiring pipelines kept running on the dead proxy without noticing.
What a Coding Interview Measures When the Candidate Has an Agent
Solo code production stopped predicting on-the-job performance once every engineer works alongside an agent. Here is what a coding interview should measure instead — and why banning or freely allowing the agent both destroy signal.
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.
