A 27B-parameter agent beat two frontier models at one task, and the task was chosen carefully

Replication is a task whose answer is already written down. The paper exists, and the agent has to arrive at its finding without being handed it ^[3]. Hughes notes this is how human scientists train, and that many PhD students begin there ^[10]. It is also gradeable, which is precisely what reinforcement learning wants, since the method rewards outcomes instead of specifying rules to follow ^[11]. Pick a task with a grader built into it and a small model can be tuned hard against that grader.

The stated ambition is bigger than that. Inherent says it wants agents that discover new scientific knowledge rather than verify old results ^[22], and it chose not to train primarily on the study of how science gets conducted, betting that reward-based training generalises across many fields ^[12]. It also set a bar above accuracy, which it calls research taste: an instinct for which experiments are worth running and how to design them ^[8]. Clearing a replication bar does not demonstrate either thing, because replication comes with a correct answer and discovery does not. Hughes describes the teammate he is aiming at as one who says "I got curious about this, and I went off and I did these experiments. What do you think of these results?" ^[24]. No replication score measures that.

What the published account does not contain is the evaluation itself: no named benchmark, no count of papers attempted, no margin over the baselines, no independent rerun ^[21]. The comparison is Inherent's, and Hughes's framing is that the method mattered more than the win ^[7].

Then there is the dependency. GPT-5.5 appears on both sides of the ledger, as a system Faraday beat and as the Codex tool Faraday used ^[19]. Inherent's analogy is that scientists lean on existing software rather than writing everything themselves ^[9]. As a cost structure it reads differently: the cheap component is the 27-billion-parameter reasoning model ^[5], and part of the working capability is rented per call from one of the vendors being outscored. Parameter count is a rough proxy for size and training cost ^[6], so an efficiency claim that stops there is measuring only the piece Inherent owns.

The hiring plan is the other thing worth pricing. Twelve people, all working in person from King's Cross ^[13], with a target of about 20 to 25 by the end of the year ^[14], which is somewhere between 1.7 and 2.1 times the current team ^[20]. At that size the scaffolding-over-scale approach is not a preference, it is the only affordable one. Hughes has joined calls to end garden leave, the UK practice of barring departing staff from joining or founding a rival for months, and says as a personal view that he was affected by it ^[16]. TechCrunch's read is that Demis Hassabis's new role has left some DeepMind staff unsettled, making Inherent a plausible landing spot ^[17].

The transferable part for anyone building agents is the sequencing. Define the job so that success can be scored without a human in the loop, then spend the budget on the reward signal rather than the model. What Faraday has not been asked to do yet is work where no answer key exists, which is the whole of the stated mission.