Why research may matter more as AI makes software cheap to build
AI made familiar software cheap to build. A startup that wants to stand out may need new knowledge, and the tools to produce it are changing: agents that run research at scale.
AI made it cheap to build software that already exists. When everyone can build familiar products, a startup that wants to stand out needs something a competitor cannot build from a prompt: new knowledge. Producing new knowledge is research.
The tools for research are changing as fast as the tools for code. Agents can now read thousands of papers, test thousands of candidates, and work in groups on a single problem.
I am both a founder and a researcher. I have written software for more than 20 years, and I am writing a thesis in Near Eastern archaeology. I think the next generation of startups will need to do research, not only ship software.
This post covers why cheap software makes research more important for startups, what automated research at scale already does, and why it has to start from what is already known.
When building gets cheap, the advantage moves elsewhere
Most new software products today are versions of products that already exist: a CRM, a website builder, a dashboard, a chat interface on top of a model.
In early February 2026, Anthropic released plugins that let its Claude Cowork agent do legal, sales, and data work. Thomson Reuters, owner of Westlaw, fell nearly 16% in a day, and RELX, owner of LexisNexis, fell 14%, Reuters reported. A price drop shows what investors fear, not what will happen. What they feared is that a general agent could do much of the routine work those companies sell.
The business professor Clayton Christensen called this pattern the conservation of attractive profits. When one step in a business becomes cheap and interchangeable, the profit usually moves to a neighboring step that is still hard.
New ideas keep getting more expensive
Producing new knowledge is not getting easier. Economists Nicholas Bloom, Charles Jones, and colleagues found that keeping chip density doubling now takes more than 18 times the research effort of the early 1970s. Ideas, in their words, are getting harder to find.
Another economist, Benjamin Jones, calls one cause the burden of knowledge. Each generation must learn more before it can contribute. A startup that wants a breakthrough faces the same cost. Automated research is the first tool in a long time that adds research effort at the price of compute.
Agents can now run research at scale
The strongest results come from agent collectives, not from a single chatbot.
Google DeepMind's AlphaEvolve runs an evolutionary loop. Models propose programs, automated evaluators score them, and the best ones seed the next round. It found a way to multiply 4×4 complex matrices with 48 multiplications, beating Strassen's 1969 algorithm, and a scheduling rule that recovers 0.7% of Google's worldwide compute.
Google's AI co-scientist splits the work between agents that generate hypotheses, critique them, rank them, and refine the winners, under a supervisor agent. It proposed how a family of bacterial genetic elements spreads between species, matching a finding that researchers at Imperial College London had made but not yet published.
Edison Scientific's Kosmos reads 1,500 papers and runs 42,000 lines of analysis code in one run, split across hundreds of agents. Lila Sciences reports that its AI-directed lab proposed, synthesized, and screened 2,942 catalysts for green hydrogen in three months. Periodic Labs raised a $300 million seed round to build labs where robots run the experiments.
Start from what is already known
In October 2025, OpenAI's Kevin Weil announced that GPT-5 had solved 10 open Erdős problems. Thomas Bloom, who runs the site that tracks them, called it "a dramatic misrepresentation". GPT-5 had found papers that already solved them, papers Bloom had not known about.
Less than three months later, OpenAI's GPT-5.2 Pro and Harmonic's Aristotle produced a formal proof of Erdős problem #728. A writeup on arXiv calls it the first Erdős problem regarded as fully resolved autonomously by an AI system.
The first episode was a literature search. The second was a new result. Automated research at scale needs both: agents that do not know the literature will spend their compute rediscovering it.
That first step is the one we built for. Agent Bayes searches a library of papers by meaning and ties each claim to the page that supports it. Our open-source research skill for Claude Code and Codex uses it to run a research project end to end: build a library of papers, check what the literature already says, run analyses, and draft a paper.
New posts, straight to your inbox
No newsletter fluff, just an email when we publish something new.
Email me when a new post is published on the Agent Bayes blog. You can unsubscribe anytime. We'll first send a confirmation email, and we only use your details for this. See our Privacy Policy.