juniper-tc02e/multi-agent-deliberation — explained in plain English
Analysis updated 2026-05-18
Get a hard question examined by multiple AI agents from different perspectives before trusting the answer.
Pick a compute tier that matches how important the decision is, from a quick quality boost to unlimited deliberation.
See a dissent register showing where agents disagreed instead of a single averaged answer.
Use the Max or Unrestricted tiers for high stakes research or design decisions.
| juniper-tc02e/multi-agent-deliberation | 0-bingwu-0/live-interpreter | 010zx00x1/faresnipe | |
|---|---|---|---|
| Stars | 2 | 2 | 2 |
| Language | Python | Python | Python |
| Setup difficulty | easy | moderate | easy |
| Complexity | 2/5 | 2/5 | 2/5 |
| Audience | developer | general | general |
Figures from each repo's GitHub metadata at analysis time.
Multi-Agent Deliberation is a set of four skills for Claude Cowork and Claude Code that let you get a question or problem examined by many AI agents working together instead of just one. Each skill breaks a problem down into smaller sub tasks, has a group of agents look at each sub task from different angles, has a separate judging step weigh their arguments instead of just counting votes, optionally challenges the conclusions to see if they hold up, and then combines everything into one final answer that also notes any disagreement along the way. The four skills, named General, Super, Max, and Unrestricted, differ mainly in how much computing effort they spend. General is meant as an everyday quality boost, using a handful of agents per sub task. Super and Max use more agents, more rounds of debate, and stricter judging, including using different AI models together rather than the same one repeated. Unrestricted has no limit at all and can keep working until it runs out of new ideas to check. The heavier tiers can spawn hundreds of agents, so picking the tier that actually fits the importance of the question matters, and each skill will tell you its plan and rough cost before it starts. The project explains that simply having several copies of the same AI model argue with each other is not automatically better, citing published research showing this can even do worse than a single model checking its own work multiple times. To avoid known problems such as debates drifting off topic, extra rounds adding cost without adding quality, and agents caving to whatever the majority says even when it is wrong, each tier builds in specific safeguards like enforcing genuinely different perspectives, stopping debates once they settle, and guarding against agents copying each other's conclusions instead of reasoning independently. Installing it in Cowork means downloading the plugin file from the releases page and uploading it. In Claude Code, you clone the repository and add it as a plugin, or build the installable bundle yourself with a provided script. No API keys, environment variables, or external services are needed. The project is released under the Apache 2.0 license.
Four Claude Cowork and Claude Code skills that break a problem into sub tasks, debate it across diverse AI agent panels, judge by argument instead of vote, and synthesize one answer with a dissent record.
Mainly Python. The stack also includes Python, Claude Code, Claude Cowork.
Free to use, modify, and distribute, including commercially, as long as you include the license notice.
Setup difficulty is rated easy, with roughly 5min to a first successful run.
Mainly developer.
This repo across BitVibe Labs
Don't trust strangers blindly. Verify against the repo.