Page 130
12 articlesAI Technical Debt: Four Categories That Never Show Up in Your Sprint Retro
Prompt rot, eval drift, embedding lock-in, and shadow coupling — four compounding forms of AI technical debt that traditional engineering practices miss, with practical strategies to manage each.
Backpressure in Agent Pipelines: When AI Generates Work Faster Than It Can Execute
Agent pipelines that spawn sub-agents and fan out tool calls create unbounded work queues that exhaust token budgets and crash production systems. Applying backpressure patterns from reactive systems — bounded queues, hierarchical budgets, circuit breakers, and adaptive concurrency — prevents runaway expansion before the invoice arrives.
Brownfield AI: Integrating LLM Features into Legacy Codebases Without a Rewrite
Practical adapter patterns — sidecar inference, async enrichment queues, and LLM-as-middleware — for shipping AI features inside legacy monoliths without a risky full rewrite.
Building Multilingual AI Products: The Quality Cliff Nobody Measures
Most AI teams ship globally with English-only evals and aggregate satisfaction scores. Here's what they're missing — and how to find the quality cliff before your users do.
The Caching Hierarchy for Agentic Workloads: Five Layers Most Teams Stop at Two
Production AI agents need five caching layers — prompt, semantic, tool result, plan, and session state — each with distinct TTLs and invalidation strategies. Most teams stop at two and leave half their savings on the table.
Capability Elicitation: Getting Models to Use What They Already Know
Most prompt optimization focuses on instruction clarity, but the real bottleneck is often the model's failure to activate knowledge it already has. A practical guide to elicitation techniques — structured decomposition, analogical priming, expertise framing — that unlock latent LLM capability without fine-tuning.
Capability Elicitation vs. Prompt Engineering: Your Model Already Knows the Answer
Most teams iterate on prompt clarity when the real bottleneck is activating knowledge the model already has. A practical guide to five elicitation techniques — from analogical priming to combinatorial prompting — that unlock latent LLM capabilities without fine-tuning.
The Centralized AI Platform Trap: Why Shared ML Teams Kill Product Velocity
Building a shared ML infrastructure team sounds like the right move. In practice, it becomes the biggest bottleneck to shipping AI features. Here's what goes wrong and what to do instead.
Chaos Engineering for AI Agents: Injecting the Failures Your Agents Will Actually Face
LLM API calls fail 1–5% of the time in production. For multi-step agents making dozens of tool calls per task, untested failure modes become customer-facing bugs. A practical guide to fault injection categories, framework design, and benchmark results for building resilient AI agents.
Consensus Protocols for Multi-Agent Decisions: What Happens When Your Agents Disagree
Majority vote among LLM agents fails nearly 24% of the time on disputed questions. Distributed systems primitives — leader election, quorum voting, and CRDTs — offer battle-tested alternatives for coordinating multi-agent decisions.
The Context Window as IDE: Why AI Coding Agents Succeed or Fail Based on What They Can See
AI coding agents fail not because models lack capability, but because retrieval pipelines load the wrong files. How context utilization, project memory files, and codebase structure determine whether your agent writes correct code or plausible nonsense.
Conway's Law for AI Systems: Your Org Chart Is Already Your Agent Architecture
Why multi-agent AI systems mirror org charts — not architecture diagrams — and the organizational patterns (embedded AI engineers, shared eval infrastructure, prompt review practices) that prevent agent boundaries from inheriting team dysfunction.
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.
