Well, 2 very intensive and educational months on every level later, here’s another summary:

  • the mentioned tools have proven themselves, additionally I’ve added opencode as an agent and quite a few MCP servers.
  • the quality differences between models are enormous. Claude models are in a league of their own (also price-wise). Through opencode I’ve gained a lot of experience with Kimi and Minimax (available for free for one context window there), which deliver quite usable results but can also go spectacularly off the rails. Due to lack of funds I’m currently mainly using the very affordable GLM4.7, which for example really struggles with Lisp-like code. I’d even go so far as to propose the thesis that LLMs have certain language preferences. Claude tends towards Python, Minimax towards bash, etc. I wouldn’t let the openrouter/free models near any relevant code.
  • MCP servers (and other plugins/tools) significantly change the vibe coding experience. To avoid having the same servers running everywhere, I can recommend an mcp-proxy.

From a developer’s perspective, LLMs are extremely powerful tools that will massively change this profession. I’m more than happy to be able to experience this exciting time, just like being there at the beginnings of the internet in the early 90s. The agent-based approach has led to a paradigm shift. I suspect it will now be about composing these agents, linking them together, etc. There are many approaches to this, one simple idea is to integrate sub-agents via ACP. Goose is currently taking this path, Hermes or Pi have been able to do this for longer, AFAIK.