Built a coding agent. Named it Sol. Bad name, acknowledged. The reasoning tracks though: coding agents are central to what HackerRank offers customers, and I'm an engineer who needed to actually understand the systems, not just ship them. Not in the "read the architecture doc" way. The "built it myself and hit every edge case" way.

There are a lot of edge cases. Tool call handling. How an agent decides when to ask versus when to act. Plugin ecosystems. CLI ergonomics. Each piece taught something. Not every lesson was deep, but each one was real. You learn differently when you're wiring it together yourself versus integrating something someone else built and documented.

The thing that crystallized hardest: you cannot YOLO anything using a coding agent. Sol is just one example. I could have pointed Opus 4.8 at a spec and one-shotted the whole system. Probably would have worked. Technically. But I'd have no idea why it worked, what the failure modes were, or where the seams were when something eventually broke. Doesn't matter if it's a coding agent, a CLI tool, an internal platform, whatever. The principle is the same. Step by step, one piece at a time, real understanding at each stage. That's not a Sol thing. That's just the rule.

There's an assumption baked into the current AI culture: more capable model plus bigger context window equals skip the learning. For throwaway tasks, maybe fine. For infrastructure you're going to maintain, extend, and explain to others, it just makes you dependent on something you don't understand. Control is the actual skill. Not a productivity metric. A prerequisite for ownership.

Not everything this week was Sol. Moved to Bangalore recently and have been working with teams here in person. Soaking up the culture, sitting in the room, understanding how people operate day to day. That's its own kind of learning. Slower, less measurable, harder to write about. Not everything needs to be about AI.