Articles
Worth reading
A short take on pieces worth your time — architecture, AI, system design, and engineering careers — each one linked back to its original source.
9 articles
Building Effective AI Agents
Anthropic's own engineers stepping back from the hype to draw a precise line between a "workflow" (LLMs and tools wired through code you control) and an "agent" (the model deciding its own next steps) — a distinction that matters the moment you're deciding how much autonomy to actually hand an LLM in production. The five patterns it lays out — prompt chaining, routing, parallelization, orchestrator-worker, evaluator-optimizer — are the ones I keep coming back to when scoping an AI feature, because they're composable primitives rather than a framework you have to buy into. Worth reading before you reach for a heavyweight agent framework: most production use cases turn out to need a workflow, not an agent.
Building Meta's GenAI Infrastructure
Meta's infrastructure team walks through the two versions of their 24,576-GPU training cluster built to train Llama and support GenAI research at a scale most of us will never operate at directly, but the design tradeoffs around networking, storage, and power show up at a tenth the size too. What I appreciated most is the honesty that GenAI workloads broke assumptions their existing infrastructure had baked in for years, forcing real architectural changes rather than just adding more machines. Good grounding for understanding what "AI infrastructure" actually means below the model layer.
Practices for Governing Agentic AI Systems
Yonadav Shavit, Sandhini Agarwal, and their OpenAI co-authors were writing about agentic AI governance back in December 2023, before most of the industry had settled on what an "agent" even meant, and the practices they propose, clear task specification, human-in-the-loop checkpoints for consequential actions, and traceability for what an agent actually did, read as more prescient than dated two years on. It's a policy paper rather than an engineering blog post, but the practices map directly onto real design decisions if you're building anything that lets a model take actions with side effects. Good context for why "just let the agent decide" is a bigger decision than it sounds.
OWASP Top 10 for LLM Applications
The LLM-specific sibling to the general OWASP Top 10, and the risks it names, prompt injection sitting at number one, plus excessive agency, insecure output handling, and supply chain risk for models and training data, are the ones I now walk through on every project that wires an LLM into something with real permissions. It's a useful corrective to how much AI security discussion still focuses on model behavior in the abstract rather than the concrete ways an LLM-powered feature gets exploited in production. Pair this with the standard OWASP Top 10 rather than treating it as a replacement; most LLM apps are still full of ordinary web vulnerabilities too.