Agents

With the advent of AI coding agents, multi-thousand-line Pull Requests are becoming the new normal. The models are churning out massive amounts of code, and keeping up is getting exponentially harder. Because I don’t have a granular understanding of every single line being merged, I’m finding it increasingly difficult to generate accurate mental models of the system’s new features. For humans, mental models are everything. We need context and an adequate knowledge of the system to reason about new architecture, confidently add features, or know exactly where to look when a production bug inevitably surfaces. ...

Stop Making Agents Write YAML At @Scale: Systems and Reliability this year, Pulumi’s Joe Duffy put a name to a problem a lot of us have been quietly routing around for years. He called it the “agentic infrastructure gap.” The setup is familiar. Andrej Karpathy has compared getting an app into production to assembling IKEA furniture: cloud consoles, API keys, configuration copied from somewhere, and a fair amount of duct tape holding it together. AI has accelerated application coding so dramatically that the bottleneck has completely shifted; writing the app is fast, but getting it running is the new hurdle. And the part that matters is that almost none of that deployment process lives anywhere an LLM can actually reason about. ...