Production agents need more visibility
I use Claude Code or Codex for almost all of my interactive coding, and Claude Code in non-interactive mode for a lot of background jobs. But the model is already nondeterministic. Putting another layer on top that I can't inspect makes it harder to understand what's happening.
I've begun experimenting with custom Pi-based harnesses around my knowledge wiki and specific code projects. Moving the context from a more mature project into Pi has started to show signs of value. For agents inside products, I've gravitated toward LangChain DeepAgents because I like its lighter base harness and the way it handles skills and tools.
It might still be too early for companies to invest a lot of time in this. I don't know where it all goes, but I'd bet that cost, control, and the ability to change models will push more teams toward purpose-built harnesses.
Originally posted on X. Lightly edited for reading here.