Page 6
12 articlesSkills Are the Package Manager for Procedural Knowledge
RAG retrieves facts, but procedures must load whole or not at all. Why agent skills behave like a package ecosystem — and why your runbooks now need versioning, testing, code review, and owners.
You Test on the Cheap Model and Ship on the Expensive One
Running CI and evals on a cheap model while production serves a frontier one breaks dev/prod parity where it hurts most. Why the tier gap invalidates your eval gates, and the budget math for closing it.
The Abstraction Layer That Made Every Model Mediocre
Multi-provider LLM gateways promise one-line model swaps but quietly strip prompt caching, schema enforcement, and reasoning controls — the features where cost and quality wins live. When the portability tax is worth paying, and the escape-hatch pattern for going provider-native.
The Accessibility Tree Is Your Newest Public API
Browser agents read your ARIA roles and DOM semantics, not your pixels — so a decade of deferred accessibility debt just became integration debt with revenue attached. How to audit your accessibility tree like the public API it now is.
The ADR Your Agent Never Read
Coding agents re-litigate every undocumented decision at machine speed. Why architecture decision records went from bureaucratic nicety to fleet infrastructure, how to wire them into agent context, and the stale-ADR failure mode nobody warns you about.
The Agent With Your Credit Card: Spend Mandates, Merchant Fraud, and Disputes When the Buyer Is Software
AI agents drove $262B in holiday orders over payment rails built for humans. How spend mandates, agent tokens, and delegation proofs work — and who eats the loss when the agent buys the wrong thing.
The Agent Your New Hire Brings: Personal AI Memory Is a Two-Way IP Boundary
Employees now arrive with personal AI assistants whose memory holds their last employer's context — and leave with yours absorbed into a tenant you can't audit or wipe. Why agent memory is a trade-secret hygiene problem, and how onboarding, offboarding, and employment contracts need to change.
The Batch Tier Is the New Spot Instance
Half your agent workload tolerates hours of delay, and providers price that tier at 50% off — yet it all runs through the interactive endpoint. A framework for classifying LLM work by latency tolerance, surviving the 24-hour batch window, and making deferral a design decision.
The Blackboard Is Back: What 1980s AI Knew About Multi-Agent Coordination
Agent teams coordinating through shared plan files are rediscovering the 1970s blackboard architecture — but they rebuilt the whiteboard and forgot the scheduler, the confidence ratings, and the level structure that made it work.
The Compiler Is the Cheapest Eval You'll Ever Run
Your language choice decides whether agents get a free, instant, deterministic verifier on every edit — or defer type errors to expensive eval suites and runtime surprises. How model authorship reorders the stack-choice tradeoffs teams settled a decade ago.
The Crash Was Load-Bearing: How LLM Tolerance Hides Broken Data Contracts
Pipelines used to enforce data contracts by crashing. LLM consumers cope with malformed input instead, turning loud incidents into silent quality decline — here's how to restore the alarm with validation gates and canary assertions.
The Fleet Forgets When the Engineer Leaves
A departing engineer's personal CLAUDE.md, custom skills, and agent memory files encode months of tacit knowledge about your codebase — and none of it survives their last day. Why agent config is the new shadow knowledge, and how to promote it into team-owned repos before it walks out the door.
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.
