Agentic Consensus Swarm — Development skill for Claude Code

Consensus Swarm

A multi-agent consensus demo — 5 agents independently answer the same question in parallel, with no visibility into each other's reasoning, and a synthesizer agent reconciles their responses into one final answer. Built with the [Claude API](https://docs.claude.com).

Live demo

**[Try it here](https://claude.ai/artifact/Usqfs4eH2g7yabwUrSHSXo)** — note: trying the interactive part requires a free Claude account to sign in with (a platform requirement for any page that calls Claude live, not specific to this project).

The problem

Hard judgment calls — should we do X or Y, what's the biggest risk here, which option is right — often benefit from more than one perspective. A single model, asked once, gives you one take. This project explores a different pattern: ask several independent "agents" the same question with different reasoning lenses, see where they actually agree versus genuinely diverge, and let a synthesizer produce a final answer grounded in that spread rather than a single unexamined take.

This mirrors, at small scale, a real pattern used in large multi-agent research systems — independent parallel attempts at the same problem, reconciled afterward — rather than one long chain of sequential reasoning (the pattern used in this portfolio's companion project, [Agent Crew](../agent-crew)).

What it does

1. 5 agents run in true parallel, each given the same question but a distinct reasoning lens — cautious/risk-first, bold/unconventional, data-and-numbers-first, customer-first, long-term-consequences-first — so they produce genuinely different attempts rather than near-identical restatements
2. Each agent's answer appears as soon as it's ready, independently, with no agent able to see another's response
3. A synthesizer agent reviews all 5 and reports, honestly, how much they actually agreed (high / mixed / low) — not just averaging them
4. A final consensus answer is produced, grounded in what the 5 agents actu