Multi-Agent Developer
A Fanboy, Not a Professional
At some point I realized I was acting like a fanboy, not a professional in AI-driven development. All I did was cheerlead Claude Code, talking about how cool they were, how the MAX plan covered everything, and so on. But in May something got very rough with payments: I had spent too much on AWS, GPU rentals, and other subscriptions. So I was not only paying $100 for Claude but also, inevitably, another $20 for ChatGPT. And at some point it hit me: something was off in my approach.
So I downgraded Claude to the $20 Pro plan. Sure, the limits ran out noticeably faster. I immediately moved all my pet projects to Codex, which I normally never used. But when I launched it, I saw essentially the same Claude — they seem to copy each other. And Codex handled a lot of not-so-important tasks just as easily.
Not every task needs your favorite tool. Sometimes you don't need the "best agent" — you need a good enough agent in the right role.
Three Tools Instead of One
The next step in my learning curve was discovering that I could also use OpenCode, and even load models into it through Ollama: Kimi-k, MinMax M3, GLM. I settled on Kimi K2.7 because it produced very good results. GLM 5.2 was released recently — I'm thinking of testing it too.
The next insight was that for all these models, including Kimi, I could still use my beloved Claude Code. I suddenly realized it's valuable not just because it has Opus inside. Claude Code itself is an agent environment, a working interface, a way to hold context, plan, and keep implementing. When the wonderful Fable 5 briefly appeared, I managed to use it literally one Friday for a single day before it was restricted, and I was thrilled: it wrote me a whole book about harness systems. I've used that book as my favorite reference ever since.
Claude Code as a Shell
But back to Claude Code as a shell for different executors. Through Ollama you can connect Kimi-k2.7-code and use all the core Claude functions, including its context. But the most interesting part turned out to be something else.
The most effective form of this coalition was: I design a task, then launch the same Kimi-k2.7-code inside Claude Code via Ollama, use the Resume command to continue a session that Opus started, and ask Kimi to continue implementing the plan. And it turned out to be very... effective. The experiment broadened my view of agents and different models instead of blindly following the hype.
Savings and Conclusion
In conclusion — significant savings. Instead of $120 for two tools, I now pay $60 for three. And the number of available models has multiplied thanks to the Ollama subscription.
I used to think AI-driven development meant finding the strongest agent and handing them the project. Now I think AI-driven development means knowing how to build a small team of models, agents, contexts, and roles.