AI Consensus circulates your prompt thru Claude, GPT and Gemini until consensus

**AI Consensus** routes a single prompt through **Claude**, **GPT**, and **Gemini**, then keeps circulating that prompt among the models until they reach consensus. The project showed up on Hacker News via the Claude/Codex/Fable channel, with the write-up at shipdiary.dev and the discussion thread at news.ycombinator.com/item?id=48995678. At listing time it had **1** point and **0** comments, so the signal is early rather than heavily validated by community debate.

Mechanically, the system treats agreement as the stop condition: one user input is fed to three frontier model families, and outputs are compared and re-prompted until the set converges. That is multi-model orchestration rather than a single-shot completion. The loop implies a controller that can pass prior answers back in, detect disagreement, and decide when answers are close enough to count as consensus. Without published latency, cost, or agreement metrics in the available summary, the architecture pitch is the product: three providers in series/parallel until they align.

For engineers and builders, the interesting part is reliability design, not novelty of any one model. When Claude, GPT, and Gemini disagree, you see model-specific bias, tool-use style, and hallucination patterns on the same task. A consensus loop can act as a cheap cross-check for code review, migration plans, security notes, or ambiguous product specs where a wrong single-model answer is expensive. It also forces you to define “same answer” in practice: exact string match, semantic similarity, or structured field equality.

In market terms, this sits next to multi-provider gateways, model routers, and agent frameworks that already call several APIs from one app. Most of those optimize for cost, latency, or fallback when a provider is down. **AI Consensus** optimizes for agreement across Claude, GPT, and Gemini instead. That is a different product bet: pay more tokens and round-trips to reduce one-model error, rather than pick the cheapest or fastest completion.

Watch whether the tool exposes how it scores consensus, caps how many circulation rounds it allows, and shows per-model drafts before the final agreed answer. Those three details decide if it is a useful engineering control or just an expensive loop. On the HN side, the **1**-point / **0**-comment snapshot means the next signal is real usage reports—especially token cost, time-to-agreement, and failure modes when the three models never converge.