Blog
- 13Apr 9, 202611 minai-agentsreinforcement-learning
How Agents Teach Themselves: The Closed-Loop Self-Improvement Architecture
A practitioner's guide to the generate-attempt-verify-train loop: how code-verifiable rewards replace human annotation, why self-play architectures double task success rates, and the three failure modes that kill closed-loop training before it pays off.
Read more → - 14Apr 9, 202611 minserverlessai-agents
The Cold Start Tax on Serverless AI Agents
Cold starts that take milliseconds for a regular Lambda function stretch to 40–120 seconds for AI agents with GPU inference. Here's the deployment decision matrix and mitigation patterns that actually work in production.
Read more → - 15Apr 9, 202610 minai-productproduct-management
The AI Feature Kill Decision: When Metrics Say Yes but Users Say No
42% of companies abandoned AI initiatives in 2025 — most waited 6+ months too long. A practical framework for recognizing when an AI feature is failing despite green dashboards, the five leading indicators that predict shutdown, and how to make the kill-or-continue decision before sunk cost psychology takes over.
Read more → - 16Apr 9, 202610 minai-productfeature-management
The AI Feature Kill Decision: When to Shut Down What Metrics Say Is Working
42% of companies scrapped AI initiatives in 2025, yet zombie features linger for months. A practical framework for recognizing when an AI feature needs to die — the behavioral signals dashboards miss, the sunk cost amplifiers unique to AI, and how to execute the kill without organizational trauma.
Read more → - 17Apr 9, 202611 minbatch-processingllm-infrastructure
The Batch LLM Pipeline Blind Spot: Offline Processing and the Queue Design Nobody Talks About
Most LLM API spend goes to batch workloads — nightly classification, data enrichment, embedding generation — yet teams design them like slow chat APIs. A practical guide to queue architecture, checkpoint-resume, failure taxonomy, and per-pipeline cost attribution for offline LLM pipelines.
Read more → - 18Apr 9, 202612 minllm-opsbatch-processing
The Batch LLM Pipeline Blind Spot: Queue Design, Checkpointing, and Cost Attribution for Offline AI
Production LLM batch pipelines fail when built with real-time serving patterns. Job sizing, checkpoint-resume, dead letter queues, cost attribution, and queue backpressure all need rethinking for offline workloads.
Read more → - 19Apr 9, 202611 mincode-agentsllm-inference
Beam Search for Code Agents: Why Greedy Generation Is a Reliability Trap
Greedy single-pass generation caps code agent reliability at 20–30% on hard tasks. Tree exploration strategies — beam search, MCTS, and structured tree search with execution feedback — deliver 30–130% pass rate improvements on the same problems without changing the underlying model.
Read more → - 20Apr 9, 202610 minllmreasoning
Cognitive Tool Scaffolding: Near-Reasoning-Model Performance Without the Price Tag
Four structured cognitive operations applied as tool calls can lift a standard 70B model from 13% to 30% on competition-level math benchmarks — nearly matching o1-preview at base-model prices. A practical decision framework for when cognitive scaffolding beats buying a reasoning model.
Read more → - 21Apr 9, 20269 minllm-latencyprompt-caching
Cold Cache, Hot Cache: Why Your LLM Latency Numbers Lie in Staging
Prompt caching makes staging latency look 80% better than production reality. A four-phase load testing methodology that accounts for cold cache, traffic diversity, and per-node routing reveals the honest p95 and p99 numbers before your users do.
Read more → - 22Apr 9, 202611 minpersonalizationllm
The Cold Start Problem in AI Personalization
When a new user sends their first message, your AI system has one data point and must make dozens of implicit decisions. Here's the architectural playbook for navigating cold start without building a filter bubble yourself.
Read more → - 23Apr 9, 20269 minmulti-agenttesting
The Composition Testing Gap: Why Your Agents Pass Every Test but Fail Together
67% of multi-agent system failures stem from inter-agent interactions, not individual defects. A practical guide to property-based invariants, trajectory replay, seam injection, and contract testing for composed agent pipelines.
Read more → - 24Apr 9, 20269 mincomputer-usegui-agents
Computer Use Agents in Production: When Pixels Replace API Calls
A production guide to computer use agents — covering the see-think-act loop, coordinate scaling pitfalls, five failure modes that kill deployments, sandboxing requirements, and a decision framework for when pixels beat API calls.
Read more →
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.
