Page 2
12 articlesThe p99 Is a Product Decision, Not an Infra One
Your tail-latency target isn't something the infra team optimizes down to a number — the number itself is a product choice. How UX affordances and latency budgets trade against each other, and why the p99 belongs in the design review.
The Semantic Diff: Reviewing a Prompt Change When the Line Diff Tells You Nothing
A three-word prompt edit can triple your hallucination rate while the line diff looks harmless. Why prompt changes need a semantic diff — comparing behavior, not text — and how to gate PRs on it.
The Token-Per-Minute Ceiling You Never Load-Tested
Provider rate limits are a per-minute token budget you don't control — and a launch is a terrible time to find the ceiling. How to forecast, load-test, and buy headroom before the 429s hit.
The Tool Belt That Grew a Long Tail Nobody Uses
Every tool you add to an agent dilutes its ability to pick the right one. The dozen tools nobody calls are quietly degrading selection on the three that matter — here is how to keep a tool catalog honest.
The User You Can't Delete: Right to Be Forgotten in AI Systems
A right-to-erasure request exposes an uncomfortable truth about AI systems: user data is smeared across model weights, vector indexes, and caches with no single row to delete. How to design for forgettability.
Chat Is the Wrong Interface: Why Your Agent Shouldn't Be a Text Box
Most people who open a chat window never send a message. Chat is a fine input primitive and a terrible operating environment — here's when to reach for structured UI, generative UI, and ambient agents instead.
Fighting the Prior: When the Model Knows a Wrong Version of Your Stack
Your coding model isn't ignorant of your framework — it knows a confident, wrong version of it. Here's why the model's prior beats your context, and the countermeasures that actually hold.
Indirect Prompt Injection: The Data Plane You Thought Was Inert
Your agent treats every document it retrieves as inert data, but retrieved text runs with system-prompt authority. How indirect prompt injection works, why EchoLeak proved it, and the defenses that actually hold.
Nobody Will Underwrite Your Agent
Your agent is accurate enough to run unattended, but deployment stalls on one question: who eats the loss when it's wrong. Why autonomy is now an insurance problem, and how to make an agent underwritable.
Pilot Purgatory: The Cheap 90% Is Why Your AI POC Can't Graduate
A working AI demo is the cheap 90% that looks like 100%. The evals, guardrails, observability, cost caps, and ownership that make it safe to run in production are the expensive 10% nobody scoped — here is a checklist that prices them up front.
Reserved Capacity for Tokens: The Reserved-Instance Decision Nobody Ported From the Cloud Era
Provisioned throughput and committed-use discounts let you reserve LLM inference the way you once reserved EC2 — but the break-even is higher than you think and your commitment can be stranded on a deprecated model. Here is the ported playbook and its one dangerous twist.
Shadow Agents: The AI Feature Legal Found Out About During the Incident Review
An unauthorized AI agent that reads customer data and takes actions is a bigger surprise than any rogue SaaS subscription. How shadow agents sneak into production, why your existing controls miss them, and the lightweight governance that surfaces them before the breach.
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.
