Faraday’s Cage Match With the AI Giants

Two. That’s the number of frontier AI labs — Anthropic and OpenAI — that a British startup called Inherent says its AI agent just beat at one of the hardest tasks in machine learning: replicating research. Not summarizing papers. Not answering trivia about them. Actually reproducing the work.

The agent is called Faraday, and Inherent — founded by DeepMind alumni — announced the result in 2026. As someone who builds bots for a living, this one made me put my coffee down.

Why Research Replication Is the Benchmark I Actually Care About

Most AI benchmarks are, frankly, easy to game. Multiple-choice science questions, coding puzzles with known solutions, chat preference scores — these measure something, but they don’t measure the thing that matters to those of us shipping agents into production: can this system take an ambiguous, long-horizon task and actually finish it?

Replicating a research paper is brutal precisely because it resists shortcuts. You have to read the paper, understand what the authors actually did (not what they said they did — anyone who has tried to reproduce a paper knows the gap there), write the code, debug it when the numbers don’t match, and iterate until they do. It’s the closest thing we have to a real-world test of agentic capability.

So when Inherent says Faraday outperformed models from Anthropic and OpenAI at this task, that’s not a vanity metric. That’s a claim about the skill that separates a chatbot from a colleague.

The “Teammate” Framing Isn’t Just Marketing

Inherent calls Faraday an AI “teammate,” and I think that word choice is doing real work. There’s a meaningful architectural difference between an assistant and a teammate.

An assistant waits for instructions and responds. A teammate holds context over time, manages its own subtasks, recovers from its own failures, and comes back to you with results rather than questions. Every bot builder I know is trying to make that jump right now, and most of us are duct-taping it together with orchestration layers, memory stores, and retry logic wrapped around general-purpose models.

If a purpose-built agent from a smaller lab can outperform the giants on a task this demanding, it suggests something I’ve suspected for a while: the frontier isn’t only about bigger base models anymore. It’s about the scaffolding, the task design, and the feedback loops you build around them. That’s territory where a focused team can genuinely compete with a trillion-dollar incumbent.

The DeepMind Diaspora Keeps Delivering

It’s also no accident who built this. Inherent’s founders come out of DeepMind, and the broader pattern is striking — DeepMind alumni have founded dozens of European startups in recent months, according to reporting on the trend. That talent pipeline is turning into a genuine wave of new labs, and Faraday is a signal that some of them are producing results that force the big players to pay attention.

For the European AI scene specifically, a British lab beating both Anthropic and OpenAI at anything is a notable moment. The center of gravity in AI has felt firmly Californian for years. Results like this complicate that story.

What I’d Want to Know Before Celebrating

Now, the builder in me has questions, because the details we have are thin:

• What was the evaluation setup? “Outperformed” can mean a lot of things. Which papers, how many, and who judged success?
• How much human steering was involved? A fully autonomous replication is a different beast from one where a researcher nudges the agent at every fork in the road.
• Does it generalize? An agent tuned hard for research replication might excel there and stumble elsewhere. That’s fine — specialization is a valid strategy — but it changes how you’d use it.

None of this is a knock on Inherent. These are the questions I’d ask of any agent, including my own. Independent verification is the whole point of replication, after all — there’s a pleasant irony in wanting to replicate the replication results.

What This Means for the Rest of Us

If you build agents, take the encouragement on offer here. The lesson from Faraday isn’t “give up, the labs will eat everything.” It’s the opposite. A focused team, a well-chosen hard task, and serious engineering can beat general-purpose systems from the biggest names in the field.

Pick a domain where finishing the job matters more than sounding smart. Build the scaffolding that lets your agent fail, notice, and retry. That’s where the next wins are coming from — and a lab named after the sc

🕒 Published: August 22, 2026