A Multi-Model Roundtable Plugin Where DeepSeek Hosts a Debate Between Claude, Qwen, and Kimi
Multi-model orchestration shifts AI from answer generation to perspective generation, but cost and coordination complexity have kept it experimental. This plugin gives a single developer a working roundtable with built-in cost controls, persistent state, and a human veto point — making multi-agent debate practical inside an existing chat harness.
Instead of getting one answer from one model, a new open-source plugin for DeepSeek Harness assembles a ring of vendor-diverse expert agents — Claude, Qwen, Kimi, MiniMax, GLM — around a DeepSeek host. One sentence in chat triggers automatic meeting creation, role assignment, and topology wiring. Experts debate with independent personas and per-round token budgets; when they deadlock, the meeting pauses and surfaces a human decision card.
The v0.2.0 release adds a visual expert management UI with real vendor logos, an on-demand knowledge base that avoids bulk token waste, per-expert output caps and opinion limits to control cost, and persistent meeting state that survives restarts. The plugin registers 13 roundtable tools on the host side, and each expert runs as a continuable sub-agent that can use a different model provider.
Under the hood, state lands in atomic JSON/JSONL files per meeting, and three token-saving mechanisms — budget circuit breakers, max-token caps, and host-relayed knowledge retrieval — keep multi-model sessions from burning through quota. The whole thing is MIT-licensed and mounts into any DSH web profile.
Treating disagreement as a first-class information source rather than a failure mode is a genuine shift from single-model answer generation — the debate structure itself produces signal.
The human decision card is the load-bearing piece: it prevents infinite loops and keeps authority explicit, which is what most multi-agent demos lack.
Token-saving here is not an afterthought — the three mechanisms (circuit breaker, per-round caps, host-relayed knowledge) are designed for multi-model economics from the start, which is rare in hobbyist agent projects.
Persisting meeting state as atomic files per meeting and decoupling it from process lifetime makes the plugin restart-resilient without a database, a pragmatic choice for a local-first tool.
Making disagreement itself into information — that concept is great. The per-round token cap plus auto-mute is very practical; multi-model meetings burn money at an astonishing rate, and having a human decision card as the fallback makes sense — machines argue, humans make the final call. Recently on ai345 I've seen quite a few discussions about multi-agent collaboration approaches; the roundtable model feels like it could be a trend this year.
Can you recommend a few that are novel? Or ones you're particularly interested in — I'll go study them.