Picture a graduate student at 2 a.m., hunched over a paper whose results refuse to reproduce. The methods section is vague, the hyperparameters are missing, and the authors have stopped answering emails. Now picture an AI agent sitting in that same chair, methodically working through the same paper, filling in the gaps, and getting the experiment to run. That second scenario is the one Inherent, a British AI lab founded by DeepMind alumni, says it has built. Its agent, named Faraday, reportedly outperformed offerings from Anthropic and OpenAI at replicating research, according to the company’s 2026 announcement covered by TechCrunch.
I want to be careful here, because the public details are thin. What we know is the claim itself: an AI “teammate” from a young lab beat the two most prominent names in the field at a specific, difficult task. What we do not yet have are the evaluation details, the task distribution, or independent verification. So rather than restate a press release, I want to explain why this particular task matters more than most benchmarks, and what it suggests about where agent architecture is heading.
Why Replication Is a Brutal Test
Research replication is, in my view, one of the most honest evaluations you can run on an agent. It resists the usual failure modes of benchmarks.
- You cannot memorize your way through it. A model that has seen the paper’s headline numbers still has to produce the pipeline that generates them. Knowing the answer does not hand you the experiment.
- It punishes shallow reasoning. Papers omit details constantly. Reproducing results means inferring unstated preprocessing steps, guessing at reasonable defaults, and recognizing when a discrepancy is your bug versus the paper’s.
- It demands long-horizon coherence. Replication is not one clever completion. It is hours of environment setup, debugging, hypothesis testing, and course correction. Most agents degrade badly over that kind of horizon, losing track of earlier decisions or looping on failed strategies.
This is why the framing of Faraday as a “teammate” rather than a tool caught my attention. A tool answers a query. A teammate holds a goal over time, notices when its own work is wrong, and comes back with something usable. If Inherent’s result holds up, the interesting claim is not raw capability but sustained, self-correcting execution.
What the DeepMind Lineage Suggests
Inherent joins a growing wave of startups founded by DeepMind alumni across Europe. That pedigree matters less as branding and more as a hint about methodology. DeepMind’s culture has historically emphasized rigorous evaluation and agent-centric research rather than pure model scaling. A team from that background choosing replication as its proving ground is a deliberate statement: they are betting that the next frontier is not larger models but agents that can operate scientific workflows end to end.
There is also a strategic read. Competing with Anthropic and OpenAI on general-purpose assistants would be a losing fight on compute alone. Competing on a narrow, high-value scientific capability is a different game, one where architecture, scaffolding, and task-specific engineering can plausibly beat scale. Whether Faraday’s edge comes from a novel base model or from superior agentic scaffolding around existing components is exactly the question I hope Inherent answers publicly.
The Questions That Decide Whether This Matters
Before anyone updates their worldview, we need answers to a few things. Which papers were in the evaluation set, and who selected them? Was the comparison against Anthropic’s and OpenAI’s raw models, or against their full agent stacks? How was “successful replication” scored, and by whom? A replication benchmark judged loosely can flatter almost any system. Judged strictly, it is one of the hardest tests in existence.
Independent evaluation is the obvious next step, and if Inherent is confident, opening Faraday to external red-teaming on held-out papers would settle the question quickly.
A Direction Worth Watching
Even with the caveats, I find this development genuinely encouraging. The field has spent years optimizing for benchmarks that measure fragments of intelligence in isolation. Replicating research measures something closer to the real job of a scientist’s collaborator. If a small lab has meaningfully advanced that capability, the implications
đź•’ Published: