Tian Pan
Software Engineer
What You Deleted Is Invisible to Your Coding Agent
Coding agents reintroduce code you deleted yesterday because absence leaves no trace in the repo. A field guide to recording the negative decisions agents need to respect.
The Nightly Batch Job That Quietly Became a Latency-Critical Service
A nightly batch job becomes a latency-critical service one reasonable request at a time. Why batch and online inference optimize for opposite goals, how the drift produces quiet failures, and how to re-architect on purpose.
The Feature Store Your Agent Reinvented Badly
AI agents re-derive the same facts every turn — churn risk, account age, plan tier — with no caching, no shared definitions, and no point-in-time correctness. Why that makes them a broken feature pipeline, and how to fix it.
Provider Rate Limits Are a Capacity Plan You Never Wrote
When your app hits a 429, the retry code that runs next quietly becomes your capacity policy. Treat rate-limit handling as deliberate load shedding — with priority tiers, jitter, and a scheduler — instead of a library default nobody reviewed.
Your Happy Path Is Your Expensive Path: The Agent That Costs More When It Wins
A failed agent run is cheap; a successful one can cost 50x more. Why raising your agent's success rate compresses margin, and the levers that fix it.
Your Agent Endpoint Is a Distributed System Pretending to Be a Function Call
An `await agent.run()` looks like a local function but hides a remote, partially-failing distributed system. Here is the timeout, retry, idempotency, and circuit-breaker discipline agent code needs.
Your Agent Has No Concept of Business Hours
AI agents act the instant they decide, with no sense of whether 3 a.m. is the wrong time to send it. How to separate anytime work from daylight work and build a timing layer that knows when to wait.
The Backfill Problem: Why Agent Memory Needs Migrations Like a Database
Agent memory is a production database that drifts every time you improve its format. Version your records, write real migrations, and backfill before old memories quietly rot.
The Bug You Can't Reproduce Because the Model Picked a Different Token
Replaying an LLM bug and watching it pass doesn't mean the bug is gone — it means you drew a different sample. How to debug a sampler when your tools assume determinism.
Build vs. Buy Is the Wrong Question for Your AI Feature
An AI feature is a stack of five layers, not one thing to make or purchase. The decision that matters is which layer compounds your differentiation and which one a competitor can simply buy.
Capacity Planning When Every Request Thinks a Different Amount
Agent requests don't have a stable cost — one resolves in 200 tokens, the next burns a million. Why p50 forecasts fail for agent workloads, and how to plan in token and tool-call distributions instead.
The Carbon Line Item Nobody Puts in the AI Feature Spec
Every AI feature review argues latency, token cost, and accuracy — and never energy. Here is how to measure per-request carbon and make it a number a team owns.