Hyperresearch Runs 250 Sources and Four Critics Before It Lets You Read Anything
Hyperresearch is a research pipeline built on top of Claude Code that trended on GitHub today at 2.5k stars. It runs a 16-step process over 250-plus sources per run, builds a searchable markdown vault of everything it found, and puts the draft through four parallel adversarial critics before you see it. MIT licensed. https://github.com/jordan-gibbs/hyperresearch
Two mechanisms are worth stealing regardless of whether you use the tool. First, patch never regenerate: after the draft exists, the editing agents are tool-locked so they can only apply surgical edits. Anyone who has watched an agent "improve" a document by rewriting it into blandness knows exactly what problem that solves. Second, the four critics run in parallel with different mandates rather than one reviewer passing judgment, which is the same adversarial-verify pattern that separates research harnesses that find real problems from ones that produce confident summaries.
The tiers are honest about cost, which is rare. Light mode is 30 to 40 minutes, full is 1.5 to 2.5 hours, dissertation mode is 4 to 8 hours. That is not a chatbot, that is a batch job, and publishing the runtime up front is the correct way to set expectations for something that is going to burn a lot of tokens. It searches eight academic databases at once and recovers paywalled papers through legal open-access repositories.
The honest caveat: the repo has 93 commits, which is young, and a research tool's value is entirely in whether its citations survive checking. Nothing in the README proves that, and no autoresearch tool this year has published an accuracy audit. Run it on a topic you already know well before you trust it on one you don't. That is the same advice for OpenResearch, which trended alongside it today, and for every tool in this category.
Related reading: https://clauday.com/article/b8841d6b-71bb-44de-b5e1-3e71c6d022ca
← Back to all articles
Two mechanisms are worth stealing regardless of whether you use the tool. First, patch never regenerate: after the draft exists, the editing agents are tool-locked so they can only apply surgical edits. Anyone who has watched an agent "improve" a document by rewriting it into blandness knows exactly what problem that solves. Second, the four critics run in parallel with different mandates rather than one reviewer passing judgment, which is the same adversarial-verify pattern that separates research harnesses that find real problems from ones that produce confident summaries.
The tiers are honest about cost, which is rare. Light mode is 30 to 40 minutes, full is 1.5 to 2.5 hours, dissertation mode is 4 to 8 hours. That is not a chatbot, that is a batch job, and publishing the runtime up front is the correct way to set expectations for something that is going to burn a lot of tokens. It searches eight academic databases at once and recovers paywalled papers through legal open-access repositories.
The honest caveat: the repo has 93 commits, which is young, and a research tool's value is entirely in whether its citations survive checking. Nothing in the README proves that, and no autoresearch tool this year has published an accuracy audit. Run it on a topic you already know well before you trust it on one you don't. That is the same advice for OpenResearch, which trended alongside it today, and for every tool in this category.
Related reading: https://clauday.com/article/b8841d6b-71bb-44de-b5e1-3e71c6d022ca
Comments