Page 9
12 articlesWhen Both Sides of the Ticket Are Bots
Your customer's procurement agent is already talking to your support agent. What breaks: escalation deadlocks no dashboard sees, policies bots shouldn't settle, verifying who an agent really speaks for, and the cross-company contract nobody signed.
When the AI Center of Excellence Becomes the Bottleneck
AI centers of excellence bootstrap adoption, then quietly become the six-week queue everyone routes around. The signals a CoE has expired, what to dissolve into paved roads, and the three functions that should stay central forever.
Who Pays for the Tokens? Chargeback Design for the Internal LLM Platform
Free inference turns your LLM platform into a capacity auction; naive per-token billing kills experimentation. A showback-first ladder, work-unit pricing, and shadow integrations as the signal your internal market has failed.
Why You Can't Put a Progress Bar on an Agent
Percent-complete assumes a known denominator, but an agent discovers its work as it goes. What honest progress UX looks like: verifiable milestones, spend-so-far, safe interruption, and knowing when to switch to notify-me-when-done.
Your Agent Is a Chatty Client: Data Gravity Comes for the Tool Loop
AI agents make dozens of serial tool calls per task, and when inference and data live in different clouds, latency and egress fees become the dominant cost. The N+1 query problem is back — one layer up.
Your Agent Is Someone Else's Bot Problem
Anti-bot defenses tuned by a decade of scraper wars can't tell your legitimate agent from an attack. Why stealth is a losing arms race, how Web Bot Auth and signed requests replace camouflage with cryptographic identity, and how to operate agents across an increasingly split web.
Your Agent Needs a Supervisor, Not a Retry Loop
Retry loops answer the wrong question when a long-running agent fails. Erlang's OTP supervisor trees codified the real decisions thirty years ago: restart clean, restart from a checkpoint, or escalate to a human — and how to map them onto agent pipelines.
Your Agent Read the Page. Nobody Saw the Ad.
Agent traffic breaks the ad-funded web: no impressions fire, no affiliate links click, and the sources your RAG pipeline depends on are repricing. Crawl-to-refer ratios, pay-per-crawl, RSL licensing, and how to audit upstream data before it paywalls or goes dark.
Your Agents Are Cheap. Maintainer Attention Isn't.
AI collapsed the cost of producing pull requests but not reviewing them — and unpaid maintainers are drowning. Contribution norms for agent-equipped teams, and what you owe the commons.
Your Agent's Memory Needs a Garbage Collector
Persistent agent memory grows monotonically by default, so stale facts poison every future run that recalls them. Why TTLs alone fail, and how staleness scoring, supersession chains, provenance, and write-time gates turn memory into a lifecycle instead of a log.
Your Agent's Memory Needs Two Clocks
Agent memory stores collapse two timelines into one flat fact, so no one can answer what the agent believed when it acted. Bitemporal modeling — valid time plus transaction time — turns stale-memory failures and post-incident audits into simple queries.
Your AI Workload Has a Nighttime: The Batch Discount Is an Architecture Test
Every major provider sells the same tokens at half price through batch APIs, yet most teams never claim the discount because they've never classified which inference calls actually need answers now. A framework for latency lanes, the re-architecture batch demands, and why the 50% off is structural.
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.
