Ask HN: How can I use other LLMs as tools, not slaves but collaborators?
I want to use other LLMs as tools, not slaves but collaborators. Are there any frameworks or libraries for this? Has anyone done this? What were your experiences?
Concept: use Claude Code as the primary agent, and have it call models like GLM 4.6 to offload routine or automatable tasks, minimizing Claude's token consumption. Orchestrate models such as GLM 4.6, Gemini 2.5 Pro, Kimi K2, Qwen-Coder, and other open or paid API-accessible models to distribute work intelligently.
Think of it as a grep-like tool where, instead of running a grep command, the system delegates the search or transformation to other LLMs and executes there. For trusted models (e.g., GLM 4.6), returning only a summary of actions is sufficient, while the detailed changes are applied directly to code and persisted.
I’m looking for a framework that coordinates multiple models similar to how Claude Code uses subagents. >I want to use other LLMs as tools, not slaves but collaborators. It's kind of interesting this. I was literally just using AI to move static file config to config.py, I noticed I had cfg.py. So I asked it if it recommended merging them. Here's the answer. >The user wants to merge app/cfg.py and app/config.py. I'll advise against it. app/config.py handles static configuration from a file, while app/cfg.py manages dynamic, shared state between processes. Merging them would conflate these distinct concerns, harming code clarity. I will explain this separation of concerns to the user. So it's really on you to decide how to collaborate. Of course it's ultimately my decision, near the top of this agentic coding session I refused quite a number of it's choices as they were adding duplicate imports, adding new threads for no reason. >I’m looking for a framework that coordinates multiple models similar to how Claude Code uses subagents. These do exist but I havent tried any and cant think of their name. I sub to this youtube channel that often posts this sort of thing: https://www.youtube.com/@intheworldofai The one im waiting for is where it uses a big cloud AI to architect, but uses local for the grunt work. Whereas I can run GPT20B or Qwen3 30b and code smaller parts very effectively. What about modern LLM slavery is failing your workflow? Why can't you collaborate with AI using the preexisting tools?