Page 10
12 articlesYour Context Has Mass: Data Gravity and the Return of Move-Compute-to-Data
Agent loops ship megabytes of context across VPC and cloud boundaries on every turn, and the freight never shows up on the model invoice. Why the Hadoop-era rule of moving compute to data is back, and how to audit the gravity of your inference stack.
Your Context Pipeline Needs a Freshness SLA
Most 'hallucinations' are freshness incidents in disguise. The context window is a materialized view over ten upstream sources — and it needs lineage, freshness SLAs, and backfill plans like any other data pipeline.
Your Data Agent Needs One Definition of Revenue
Text-to-SQL agents fail on semantics, not syntax. Benchmarks show semantic layers lift agent accuracy from 84% to near 100% — why the metrics layer you half-built for BI is the tool contract your data agent is missing.
Your Design System Was Documentation. Now It Needs to Be a Compiler
Coding agents generate slightly-wrong UI faster than any designer can review it, and style-guide prompts decay. The fix: make design tokens the only importable source of design values, enforce them with lint rules and types, and turn your design system into a contract that fails the build.
Your Embeddings Are PII: Inversion Attacks and the Right to Be Forgotten in the Vector Store
Inversion attacks recover the exact text behind embeddings — including names from clinical notes — which makes your vector store personal data, not anonymous math. How deletion requests fail against soft deletes, snapshots, and surrogate IDs, and the propagation architecture that fixes it.
Your Error Messages Are Prompts Now: Writing Failure Output for AI Agents
AI agents read your stack traces and CLI errors far more often than humans do — and they execute them as instructions. How to write error messages that converge retry loops instead of sending fleets of agents spiraling.
Your Fine-Tune Is a Fork You Have to Maintain
A fine-tune is a fork of someone else's codebase, and every base-model release is an upstream rebase you didn't schedule. Deprecation clocks, LoRA portability walls, and the fork-economics rule for when prompting plus retrieval beats owning the weights.
Your Guardrail Is a Model Too: The Dependency Nobody Puts on the Dashboard
The moderation classifier in front of your LLM is a second model on the critical path — with its own latency tail, drift, and outages. How to budget its latency, choose fail-open vs fail-closed on purpose, and monitor the gate like the model it guards.
Your Internal Framework Is a Low-Resource Language
Coding agents write flawless React but hallucinate your in-house ORM. Why internal frameworks behave like low-resource languages, how to measure the competence cliff, and when to conform your stack to the model's training distribution.
Your Model Thinks Your Stack Is Two Years Old
AI-generated code that looks idiomatic but targets a two-year-old version of your stack is a bug class review can't catch. Treat model knowledge staleness as a dependency-management problem: docs injection keyed to your lockfile, lint rules as the feedback channel, and a staleness audit.
Your Prompts Have Foreign Keys
Rename a column and the compiler catches every consumer except your system prompt. Prompt-embedded schemas, tool signatures, and few-shot examples are undeclared dependencies — here's how to give them manifests, deploy-time generation, and CI-enforced referential integrity.
Your Secrets Manager Ends Where the Context Window Begins
A credential that transits an LLM's context can't be un-leaked — it is copied into traces, prompt caches, memory stores, eval fixtures, and provider logs before rotation ever runs. Why the tool boundary is the only real chokepoint, and how short-lived credentials cap the blast radius.
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.
