Scientific replication is the grind work of research: parse a methodology, reconstruct the inputs, execute the procedure, check the output against the original results. It is also the step that most AI agents cannot do reliably. British AI lab Inherent, founded by DeepMind alumni, released an agent called Faraday that it says clears the full replication sequence and outperformed systems from Anthropic and OpenAI in doing so.
The replication problem
Execution fidelity across a full methodological chain is where automated scientific work breaks. An agent can read a paper and still fail to reproduce it.
Research methodologies are written for human readers who carry implicit domain knowledge. A phrase like "standard conditions" or "as previously described" assumes the reader knows what that resolves to in context. An agent has to resolve those ambiguities explicitly, and each resolution is a branch point where errors compound forward. A misread variable range or a wrong baseline assumption propagates through every downstream step. The longer the methodology, the wider the divergence window becomes by the time the agent reaches the output.
Faraday is built specifically for that constraint. Inherent describes the agent as a "teammate" for researchers, a framing that places it inside the experimental workflow rather than as a summarization layer above the literature. Where this sits in the research stack is at the execution layer: not explaining what a method says, but running it.
What Inherent is claiming
Inherent says Faraday outperformed agents from Anthropic and OpenAI on research replication tasks. That comparison comes from Inherent's own characterization of its release; the available material does not cite an independent benchmark methodology or external verification. The lab is British, founded by alumni of DeepMind, which has published extensively on agent architecture and scientific AI. That lineage gives the claim context without validating the comparison.
Inherent says Faraday's replication capability could be a stepping stone for scientific innovation. An agent that can take a published methodology and reproduce it reliably shortens the cycle between a paper's release and the experiment that builds on it. The benchmark methodology is not publicly available for independent review.