Could it be tooling like Claude Code? I just used Claude Code with qwen3.5:35b running locally to track down two obscure bugs in new Common Lisp code I wrote yesterday.
back
2 comments
I use Claude Code as an orchestrator and have the agents use different models:
product-designer ollama-cloud / qwen3.5:cloud
pm ollama-cloud / glm-5:cloud
test-writer claude-code / Sonnet 4.6
backend-builder claude-code / Opus 4.6
frontend-builder claude-code / Opus 4.6
code-reviewer codex-cli / gpt-5.1-codex-mini
git-committer ollama-cloud / minimax-m2.5:cloud
I do in similar way, connect claude code to litellm router that dispatches model requests to different providers: bedrock, openai, gemini, openrouter and ollama for opensource models. I have special slash command and script that collect information about session, project and observed problems to evaluation dataset. I can re-evaluate prompts and find models that do a job in particular agent faster/cheaper, or use automated prompt optimization to eliminate problems.
Is this because Anthropic models are worse at those tasks, or more expensive, or what?
They are great, but they are expensive. I can run those against the cheaper ollama cloud models for things that are basically requirements gathering and review of a plan. The Product Designer Agent and the Product Manager basically argue for a few rounds and give an artifact that the coding agents pick up.
It could all easily be anthropic models and would work well, but running this swarm eats up all my anthropic tokens and these other models are good enough for the roles I've given them.
very cool!
Qwen seems fine for analysis to me, but Opus 4.6 is far better to use as a sounding board or for writing code