Orchestrating Agentic Coding

I had a conversation with a friend recently that crystalized something I’ve been thinking about for months. He was frustrated with agentic coding tools - expected them to excel at straightforward tasks like reorganizing files and fixing imports, but instead found them constantly overengineering solutions and losing focus on the actual problem.

This perfectly captures the current state of AI coding tools. They’re simultaneously impressive and frustrating, powerful and unpredictable. But here’s the thing: I think we’re approaching an inflection point where the real innovation won’t be in making these tools smarter, but in orchestrating them better.

Forget the stochastic parrot nonsense. These tools are getting close enough to “good enough” that with proper validation, we can consistently rely on them to generate workable products. The key insight isn’t that they’re perfect - it’s that they’re predictably imperfect in ways we can work around.

Think about it like this: if you can spin up multiple AI coding instances, let each tackle the same problem from different angles, and then use deterministic testing to validate the outputs, you’ve essentially created a parallel development pipeline where quality emerges from competition rather than individual genius.

Leave a Comment