Page 5
12 articlesGraphQL Finally Found Its Client, and It Isn't Human
AI agents are the first clients that actually compose their own data needs — the client GraphQL was designed for. But agents also amplify every weakness that made GraphQL lose the human-client war. The pattern that works: agents compose queries at dev time, humans review and pin them as persisted operations.
Headcount Planning When Compute Writes the Code
Annual planning still converts roadmap ambition into requisitions, but in an agent-heavy org the marginal unit of output is tokens plus review bandwidth. How to size absorption capacity, budget inference, and write a plan where hiring and compute compete for the same outcome.
Let the Agent Answer the Page First: A Trust Ladder for AI Incident Response
The first fifteen minutes of an incident are mechanical, and an agent can do them before your laptop unlocks. A four-rung trust ladder — narrator, proposer, gated executor, bounded auto-remediator — with evidence-based promotion and scoped demotion beats banning automation after one bad night.
No-Code Was a Bet That Code Stays Expensive. The Bet Just Lost
No-code platforms priced themselves against expensive engineers. Now that AI agents build and maintain real apps for less than a seat license, the abstraction is a ceiling — and the survivors are pivoting to selling governed data access instead.
Notebooks Are Hostile Territory for Coding Agents
Hidden kernel state breaks every assumption a coding agent makes: the file isn't the program, outputs lie, and re-running cells is unsafe. Why restart-and-run-all is the only contract an agent can verify, how pure cells make notebooks agent-safe, and the tooling notebooks now owe their newest users.
Performance Reviews Measure the Fleet, Not the Engineer
PR counts and velocity now measure your agent harness, not your engineering judgment. How calibration, metrics, and career ladders must change when AI writes half the diff.
Point-in-Time Restore for Systems That Learn: The Backup Nobody Takes
Your database has snapshots and your code has git, but your agent's brain is smeared across five stores that version independently. Why 'restore to yesterday 3pm' is undefined for most agent stacks, and how storage engineering's consistency-group discipline fixes it.
Ports and Adapters for Agents: Why Your Tool Schemas Should Outlive Your Provider
Provider migrations break at the tool layer, not the prompt. Treat tool schemas as ports you own — with thin adapters per provider and per backend — and deterministic agent tests come free.
Postel's Law Is a Liability at the Tool Boundary
Lenient parsers that coerce near-miss tool calls teach AI agents that sloppy calls work, compounding errors across the loop. Strict validation with legible, example-bearing errors is in-loop training signal — here's when to be strict and where Postel's law still wins.
Sign Your Weights: Your Model Is an Executable Your Supply Chain Ignores
Your CI signs containers and pins package hashes, but the model weights deciding your product's behavior load unverified. How pickle backdoors, poisoned checkpoints, and open buckets became the unguarded attack surface — and how OMS signing, safetensors, and digest pinning close it.
When the Hardware Lies: Silent Data Corruption Meets Stochastic Software
Roughly one accelerator in a thousand silently computes wrong answers, and LLM inference is the first major workload where hardware faults and sampling noise look identical. How bit flips hide inside stochastic outputs, and the canary lanes and per-host statistics that catch a lying GPU.
Single-Use Software: Write It, Run It, Delete It
AI made one-off scripts cheaper than finding and configuring existing tools — but ephemeral code skips security review and erases institutional learning. What disposable software really costs, and the four rules for when a throwaway must become a product.
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.
