Orchestrating local LLMs: assign skills per model, let a third one debate and verify
gnukeith · x · 2026-11-17
The author describes a homegrown orchestration skill: when multiple local models run together (e.g., Qwen, DeepSeek, GLM Flash), a set of rules assigns different skill sets to each—one handles deep research and documentation, another physics/science/math, and a third triple-checks the others' work by debating them until they agree, or helps finish the report.
He admits the setup isn't optimal: it's extremely token-hungry and cheaper or more thorough alternatives surely exist, but he values the rigor and finds it good enough.
More from coding & agent
- Voice agent passes 15 metrics but customers still complain — what's not being measured? — georgeamine · 2026-11-17
- Google's modern-web-guidance lifts coding agents' best-practice pass rate from 52% to 92% — victor_explore · 2026-11-17
- Open-source Irrigator adds crash recovery and observability to coding agent swarms — volatilityfund · 2026-11-17
- 435K LOC deep dive: a hands-on manual for LangChain 1.5, LangGraph 1.2 and Deep Agents — ghumare64 · 2026-11-17
- rauchjulian: traces are the essential new primitive of software development — your intelligence — HamelHusain · 2026-11-17
- Developer Lets an AI Agent Play GTA5 Live to See How Far It Gets — zhuowei · 2026-11-17