DeepMind Alumni Build AI Agent That Outperforms OpenAI and Anthropic at Research Replication

---

DeepMind Alumni Build AI Agent That Just Beat OpenAI and Anthropic at Replicating Research

A British AI lab founded by veterans of Google DeepMind has released an AI agent that claims to outperform models from both Anthropic and OpenAI on one of the most demanding tasks in science: replicating published research papers. The implications extend well beyond the laboratory.

Inherent, the startup behind the agent, announced Faraday this week — a purpose-built AI "teammate" designed to replicate scientific experiments from academic papers. According to the company, Faraday's benchmark results surpass those of competing models from two of the most well-resourced AI labs in the world. The story was first reported by Anna Heim at TechCrunch AI.

What Faraday Actually Does

Research replication is not a flashy use case, but it is a critically important one. In science, the ability to reproduce results from a published paper is considered the gold standard for verifying that a finding is real, not a fluke or a fabrication. The so-called "replication crisis" has undermined trust in large bodies of research across fields from psychology to medicine.

Faraday is designed to take a scientific paper and reconstruct the experiments it describes — the methodology, the code, the analysis — with enough fidelity that the results can be independently verified. If an AI agent can do this reliably, it could accelerate the pace at which new research is validated, built upon, or discarded.

That Faraday is reportedly outperforming models from OpenAI and Anthropic at this task is a meaningful signal. It suggests that specialized, domain-focused AI agents — rather than general-purpose large language models — may be better suited to high-stakes, structured tasks that require precision over creativity.

The Bigger Picture: Specialized AI Agents Are Gaining Ground

The release of Faraday is part of a broader trend that business leaders should be paying close attention to. General-purpose AI models have dominated headlines for the past few years, but the emerging competitive edge is increasingly found in agents built for specific workflows.

Faraday is not trying to write poetry or summarize emails. It is engineered to do one thing well: take structured scientific knowledge and reproduce it accurately. That narrow focus appears to be paying off in measurable performance gains over much larger, more expensive models.

This pattern is showing up across industries. Specialized agents are beginning to outperform generalist models on legal document review, financial analysis, compliance checking, and now scientific research. For business teams evaluating AI tools, this is a critical distinction. Deploying a general-purpose AI for a specific, high-stakes workflow is increasingly not the optimal choice.

What This Means for Business Teams

For most companies, the immediate application of a scientific research replication tool is limited. But the principles behind Faraday's design are directly applicable to business operations.

Consider the parallel in corporate knowledge work. Many organizations struggle with a version of the replication problem: onboarding new employees, recreating past analyses, auditing decisions, or verifying that internal processes are actually being followed as documented. These are workflow problems, not science problems, but they share the same underlying challenge — can you take a documented process and reproduce it reliably?

The answer from specialized AI agents is increasingly yes. And the companies that recognize this earliest will compound significant advantages in consistency, speed, and quality of output.

Small and mid-sized businesses in particular have the most to gain here. Large enterprises have armies of analysts and researchers. SMBs typically do not. An AI agent that can handle the methodical, structured work of replicating, verifying, and executing documented processes could be a genuine equalizer.

For teams already exploring how to integrate AI into their daily workflows, tools like AI tools for business are worth examining in light of this shift toward specialization. Understanding the difference between a generalist model and a purpose-built agent — and when to use each — is quickly becoming a core operational skill.

The DeepMind Factor

It is also worth noting who built this. DeepMind has long operated at the frontier of applied AI, producing breakthrough work in protein folding, game-playing, and scientific modeling. The alumni who founded Inherent are not newcomers to the challenge of building AI systems that perform at benchmark-level precision. That pedigree lends credibility to the company's performance claims in a way that a typical startup announcement might not.

As AI agents become more capable, platforms that help teams evaluate, deploy, and manage those agents will become essential infrastructure. WRRK.ai is built for exactly this moment — helping business teams cut through the noise and put the right AI tools to work.

---

Explore how AI agents are changing team workflows and stay ahead of what is coming next.

Original reporting by Anna Heim, TechCrunch AI, published August 22, 2026. Read the original article at TechCrunch.

Start putting smarter AI agents to work for your team at WRRK.ai.

---

Frequently Asked Questions

What is Faraday and who built it?

Faraday is an AI agent developed by Inherent, a British AI startup founded by alumni of Google DeepMind. It is designed specifically to replicate scientific research papers — reconstructing the methodology, code, and analysis from published studies with high accuracy. According to Inherent, Faraday outperforms models from both OpenAI and Anthropic on research replication benchmarks.

Why does AI research replication matter for businesses?

Research replication is about verifying that documented processes and findings can be reliably reproduced. In a business context, this maps directly to workflow consistency, process auditing, and knowledge transfer. AI agents capable of executing structured, documented tasks with high fidelity could dramatically reduce the time and cost of onboarding, compliance, and operational quality control — particularly for small and mid-sized businesses without large support teams.

Are specialized AI agents better than general-purpose models like ChatGPT?

For specific, high-stakes workflows, the evidence is growing that specialized agents outperform general-purpose models. Faraday's performance against OpenAI and Anthropic on research replication is one example. General-purpose models excel at breadth and flexibility, while purpose-built agents are optimized for precision in a defined domain. Business teams benefit from understanding this distinction when selecting AI tools for particular use cases.