Page 149
12 articlesAlphaEvolve's Architecture: How Evolutionary Search + LLMs Discovered a Better Matrix Algorithm
A breakdown of AlphaEvolve's four-component loop — program database, prompt sampler, LLM ensemble, and evaluator — and what engineers can learn from the architecture that beat a 56-year-old algorithm.
Evaluating AI Agents: Why Grading Outcomes Alone Will Lie to You
A practical guide to evaluating AI agents by grading both outcomes and multi-step trajectories — covering grader types, pass@k vs pass^k, eval harness design, and the organizational pitfalls that sink evaluation programs.
Context Engineering: The Invisible Architecture of Production AI Agents
Context rot degrades every major LLM at scale. Learn how to manage context as first-class infrastructure—KV-cache optimization, reversible compression, error trace retention, and the metrics that reveal degradation before your first production incident.
Agentic Engineering Patterns: The While Loop Is the Easy Part
Every production agent runs the same trivial loop. The patterns that matter are the ones built around it — prompt chaining, routing, reflection, and the context discipline that prevents $47,000 weekly bills.
Revisiting Trade-offs: Think Like a Fox, or Focus Like a Hedgehog?
Using the Hedgehog Concept to discover core competency in startups. Learn how to identify the intersection of passion, capability, and economic engine to avoid distraction and achieve long-term value.
Building a Multi-Agent Research System: Patterns from Production
How multi-agent research systems actually get built — the architectural patterns that work, the failure modes that bite in production, and the engineering discipline required to keep costs and quality under control.
Governing Agentic AI Systems: What Changes When Your AI Can Act
When AI agents can call APIs, write to databases, and spawn sub-agents, governance shifts from controlling outputs to controlling actions. A practical engineering framework for authorization, minimal footprint, prompt injection defense, and structured human oversight.
Agent Engineering Is a Discipline, Not a Vibe
Why most AI agents fail in production — and the six structural dimensions (intent, memory, planning, control flow, authority, tools) that separate reliable systems from ones that only work in demos.
Designing an Agent Runtime from First Principles
A production agent runtime is not a function runner — it is an execution substrate. Here is how to design one correctly, covering graph execution models, checkpointing, human-in-the-loop, and observability from first principles.
Why Your Agent Should Write Code, Not JSON
Code-action agents let LLMs emit and run Python instead of JSON — achieving 20% higher task success rates and 30% fewer LLM round-trips. Here's how they work, where they fail, and how to run them safely in production.
Why Your AI Agent Wastes Most of Its Context Window on Tools
Every tool definition your agent loads is a token tax paid upfront. With 50+ MCP tools connected, definitions alone can consume 130K tokens before any work begins. Here are the three bottlenecks breaking production tool use and the patterns that fix them.
Why Your AI Agent Should Write Code Instead of Calling Tools
Code-executing agents can cut token usage by 98–99% compared to standard tool-calling patterns — and that's just the start. Here's how the architecture works, where it breaks, and when to use it.
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.
