Page 3
12 articlesThe AI Roadmap Priced In a Model That Doesn't Exist Yet
Scoping features around where models will be in six months turns a plan into a forecast on a vendor's release schedule. How to tell a defensible capability bet from wishful thinking, and design so a better model is upside, not oxygen.
The Conversation With No Owner: Accountability Dissolves Across an Agent Handoff Chain
In multi-agent systems, every agent can pass its own test while the whole conversation fails. Why handoffs diffuse ownership, and the primitives that put accountability back.
The Eval That Scored Every Turn and Missed the Whole Conversation
Turn-level accuracy of 95% does not mean 95% of sessions succeed. Here is why averaging per-turn scores hides compounding failures, and how to evaluate whole conversations instead.
The Guardrail Tax: When Your Safety Classifier Doubles Your Latency and Your Bill
Bolt-on safety classifiers stack in the critical path, tripling latency and inflating your inference bill. How to size guardrails to blast radius and run them concurrently instead of serially.
The Model Upgrade That Broke Nothing the Tests Could See
You move to the newer model, every eval passes, latency drops, and support tickets climb anyway. Here is the class of regression that passes every assertion, and how to catch it before your users do.
The Noisy Neighbor Is You: When Runaway Agents 429 Everyone Else on the Shared Account
A shared model account has one org-wide token budget, and any unpaced batch job can starve every other team's production traffic. Here is why per-org limits make agents a shared-fate resource, why backoff alone stampedes, and the isolation patterns that hold.
The Postmortem With No Root Cause
Five-whys incident review assumes a deterministic cause the model doesn't have. Here is how to write an honest LLM postmortem built on failure-rate deltas and contributing-factor stacks instead of a single root cause.
The Prompt-Cache Cliff: How One System-Prompt Edit Re-Priced Your Whole Fleet
A one-line edit to the top of your system prompt can invalidate every cached prefix at once, dropping your cache hit rate to zero and multiplying your inference bill overnight — here's why, and how to structure prompts so it can't happen.
The Staging Environment That Has No Staging Model
A hosted model is the one dependency you can't stand up a faithful staging copy of. Here's why pre-production parity breaks for LLMs, and how pinning, replay, golden transcripts, and shadow traffic each close only part of the gap.
Two Model Vendors, One Feature: The Redundancy That Buys a Consistency Nightmare
Adding a backup LLM vendor feels like textbook resilience, but two models are not two replicas. Here is the hidden consistency and maintenance tax nobody budgets for.
Your Agent Transcripts Are Discoverable, and Legal Never Signed Off
Your agents log every reasoning step because observability begged for it and storage is cheap. The same trace that helps you debug is a subpoena magnet. Keep the action ledger, expire the deliberation, and redact the PII before it lands.
Your AI Bill Is One Untagged Line Item: FinOps When Tokens Refuse to Be Tagged
Model APIs bill by key and by time, not by feature or customer, so your AI spend lands as one untagged line item. Here is why token cost resists cloud-style tags and what to instrument at call time to fix it.
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.
