Page 7
12 articlesThe Internal Capacity Market: Rationing Scarce Inference Between Teams
One team's eval sweep can starve another team's production chat when both draw from a shared inference pool. Borrowing from mainframe time-sharing and Borg's priority bands: priority tiers, preemption, chargeback, and when a team deserves dedicated capacity.
The LLM Contract Clauses That Actually Matter: A Negotiation Checklist for Enterprise AI Buyers
Model deprecation notice, zero data retention, copyright indemnity exclusions, and contractual rate limits — the LLM contract clauses that actually move risk for enterprise AI buyers, and which levers vendors will genuinely negotiate.
The Merge Queue Is the New Bottleneck
Coding agents doubled your PR volume, but merges still crawl through a queue sized for human typing speed. The capacity math behind merge-queue congestion, why flaky tests cascade at agent scale, and how tiered pre-merge/post-merge testing restores throughput.
The Model API Is Tier 0 Now. Design the Degraded Mode Before the Status Page Turns Red
Model APIs sit in the request path next to your database, but most DR plans still treat them as nice-to-haves. A practical guide to degraded modes — caches, predecessors, queues, and honest downtime — and the business decisions to make before the next provider outage.
The Model Is Already Talking to Your Customers
AI assistants quote your old pricing, recommend dead endpoints, and pitch competitors in conversations you never see. How to monitor model answers like a production surface and ship ground truth machines can read.
The Morning Review Queue: Triaging Eight Hours of Unattended Agent Work
Overnight coding agents don't hand you finished code — they hand you a triage problem. How trust tiers, provenance receipts, batched decisions, and two-pass triage keep the morning review queue from becoming your team's biggest time sink.
The p99 of Thought: When the Model Decides How Long Your Request Takes
Reasoning models make response time a function of problem difficulty, so p99 latency spikes 3-5x above p50 and timeout tuning turns into a cost problem. Hedged requests against smaller models, per-route effort caps, and difficulty-aware SLOs keep the tail under control.
The Package Your Agent Hallucinated Now Exists — and It's Malicious
LLMs hallucinate the same fake package names again and again — 43% recur on every rerun — and attackers register them. Why agents with install rights turn a hallucination into a supply-chain compromise, why diff review can't catch it, and the lockfile, allowlist, and cooldown defenses that can.
The Real Backend Is a Spreadsheet — and Your Agent Just Got Write Access
Excel and Google Sheets hold most companies' operational logic — untyped, untested, unreviewed — and AI agents now write to them at machine speed. Protected ranges, named ranges, formula diffs, and migration-style writes keep one overwritten formula from becoming a business decision.
The Rewrite Is Cheap Now. Being Right Isn't.
Coding agents collapsed the cost of translating a legacy codebase — but undocumented behavior, data migration, and integration cutover cost what they always did. A framework for which rewrites the cheap-translation era actually unlocks.
The RFC Glut: What Happens to Design Review When Writing Is Free
AI agents can draft a polished eight-page RFC in an hour, killing the effort signal design docs once carried. How to reprice design review: live defense, pre-registered metrics, prototypes over prose, and explicit review budgets.
The Rollback That Couldn't: Prompts, Tools, and Memory Version Together or Not at All
Rolling back an agent's prompt while its tools and memory advance doesn't restore a known-good state — it ships an untested chimera. Why agent deployments are a triple, how each one-legged rollback fails, and how to pin prompts, tool contracts, and memory like a lockfile.
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.
