Research and development organizations in 2026 find themselves in a unique position. AI agents are increasingly implemented into business processes. Coordinating agents are delegating fold predictions to specialist sub-agents.
Reviewer agents check the citations. Scientists look at the results and move on. Yet no one signs off for the record because a signature was never requested. This gap between the output and true human approval has become a tricky gray area.
Claude Science is one of the catalysts for why this question is now at the forefront of many engineers’ minds. Launched in June 2026, it’s an AI workbench that signals agentic AI is now moving from a pilot phase to actual implementation in the scientific R&D process.
Its release also aligns with two regulatory developments that have been gaining attention. One is the draft European Union annex drawing a line around which GMP-critical work an AI agent could touch. The other is a U.S. Food and Drug Administration warning letter flat out demanding human signatures on AI work.
Taking these three events into account, one practical question emerges: Who should be accountable for the outputs of an agentic AI agent? Kivo, a provider of AI compliance solutions, has set out to answer this question by digging into the data on what Claude Science is, along with an analysis of the two regulatory announcements, so that R&D teams can put the right people in place for sign-offs.
Understanding why there’s so much attention on the agentic AI space specifically in an R&D setting requires taking a look at why Claude Science, and other similar technologies, are making waves. Released by Anthropic, Claude Science was designed to be a single environment that folds together research tools that scientists normally juggle across multiple browsers and terminal windows.
It comes pre-configured with over 60 databases and connectors that span genomics, single-cell analysis, proteomics, structural biology, and chemical information. Essentially it’s one cohesive system. What makes it so impactful is its native connection to NVIDIA’s BioNeMo models, which include Evo 2, Boltz-2, and OpenFold3.
Perhaps more notable is the fact that Claude Science is an expansion of their Claude for Life Sciences product, which has already integrated Claude to platforms like Benchling, PubMed, and 10x Genomics.
Its workflow runs on a multi-agent structure. A generalist coordinating agent addresses a researcher’s immediate request. It then determines sub-specialists best suited for the tasks. Anything from parsing a genomics pipeline to folding a protein could be fair play.
A separate reviewer agent is then pulled in to look at the work and flag anything incorrect, whether it’s improper citations, data that can’t be traced back to source code, or something else. Then, a final answer is spit out to the researcher.
In their launch post for the product, Anthropic touted how, in its beta, the technology was used for single-cell RNA sequencing analysis, CRISPR screen design, protein structure prediction, and even chemoinformatics work. All impressive feats.
Its success led Anthropic to announce that they will provide funding for up to 50 AI for Science projects running from September through December 2026. This means a substantial number of labs are going to be producing many agent-assisted results in a short window. This is what has some experts concerned.
The crux of the issue comes back to Anthropic’s pitch and their claim of auditability. The company has noted that every figure produced by Claude Science comes with the exact code that generated it. This includes plain-language descriptions of how the data was created and its history. In practice, this means a researcher could open it up months later to understand how and why the data was produced.
It’s an astounding capability that addresses the long-standing problem of being able to trace AI outputs in layman’s terms.
However, auditable and regulatory compliant are not always the same thing. It’s one thing to read a code and determine the logic an AI system used to draft an answer, but it’s another to show who was accountable for signing off on the data or whether that data is even still accurate. This creates ambiguity related to the 21 CFR Part 11 regulation, which has governed electronic recordkeeping since the 1990s.
This regulation stipulates that for something to have a true audit trail, it must have a secure, time-generated computer stamp and be tied to an authenticated human identity. It’s required for the data to count as an official record. For many experts, Anthropic's Claude Science builds a strong reproducibility claim, but whether its audit trail is in compliance with Part 11 is a little more unclear.
The release of Claude Science coincides with two global regulatory issues. The EU and FDA have both raised concerns around AI, each of which has overlaps with the way Claude Science operates.
Draft EU GMP Annex 22
As detailed in the European Commission's 2025 draft, Annex 22 draws some of the sharpest lines yet between static and adapting AI models. The latter is excluded entirely from GMP-critical applications on the basis that if a model keeps changing its behavior and learning over time, it is an uncontrolled variable inside of a set system.
Claude Science and other similar workbenches fall square within these sights. They’d be permitted in noncritical functions, but there would need to be documented human oversight and defined areas where a real person can get in and intervene.
The Annex 22 draft also closes a loophole that’s been around for some time. The vendor who issued the model is the one responsible for GMP compliance. The resolution isn’t binding yet as it’s still under revision, but the final version is expected to hit at the end of 2026.
FDA's April 2026 warning letter — human review as the bright line
Annex 22 addresses models themselves, but the FDA took aim of the workflow itself in a warning letter to Purolea Cosmetics Lab on April 2. In it, the agency outlined that they found the lab had used AI agents to draft drug specifications, procedures, and master production records, then used those records without further human review. This was alleged as a direct violation of the CGMP requirements at 21 CFR 211.22(c).
In one specific response, the FDA signaled the company had indicated they skipped a required validation step solely because the AI agent that was used hadn’t ever indicated the step was necessary.
An industry analysis of the warning letter in ECA Foundation outlined the ramifications. All results or recommendations from an AI system need to be reviewed and approved by a human in a company’s quality control unit before it becomes a controlled record. This falls under Section 501(a)(2)(B) of the Federal Food, Drug and Cosmetic Act.
The reason this made waves is because it serves as the first formal FDA warning letter addressing AI agent use directly and sets the precedent that authorized human review is now something inspectors can cite.
Companies aren’t just sitting idly by. Regulatory direction has started to become exceedingly clear and, with workbenches already in scientists’ hands, life science software vendors are working to build connective tissue between technologies.
AI technology is still in its infancy. As more tools and products continue to be rolled out, more and more questions will inevitably pop up. Here are some common ones being asked of IT and quality teams:
If the coordinating agent delegates to a sub-agent, whose output is the record — the sub-agent's, or the coordinator's summary of it?
In most workbenches, the record is what the coordinator spits out to the researcher. This means that much of the sub-agent’s work can go untracked, including any judgement calls, unless the platform it’s using is specifically built to avoid this. Quality teams reviewing agentic tools should start asking vendors to show full delegation chains and not just a single output.
What happens when the model version updates mid-project? Do we need to re-validate every analysis run on the old version?
Annex 22 gets at this issue a little bit. Under a static vs. adaptive reading, a model version change is a change-controlled event. The safest practice is to just pin a single model version to a study or submission for its entire duration. If there was a change midway through, log which versions produced which results in the same way a lab treats an instrument switch in the middle of a study. Proper documentation is key to compliance.
If the reviewer agent flags and 'fixes' an error before a human ever sees it, is that fix part of the auditable history, or does it disappear into the final output?
A correction to data that a human never saw is still an AI judgment call that shaped the auditable record. For that reason, it needs to be documented.
Our data governance policy prohibits certain datasets from leaving our environment — does calling out to BioNeMo/Modal compute count as “leaving”?
The answer to this will be specific and dependent on the architecture an organization is using. Some workbenches like Claude Science are built to keep sensitive datasets on infrastructures a lab already controls, sending out only context that’s needed. This distinction in how the models handle data is exactly what needs to be spelt out in internal governance policies.
Under Annex 22 logic, is a CRISPR screen design a 'critical' GMP application or a research-stage, noncritical one — and who decides that line?
The draft language for Annex 22 ties criticality to whether a model’s output is feeding a decision that affects product quality, design, safety, or integrity. It has nothing to do with the sophistication of the science. Early-stage CRISPR screen design will read as noncritical for most organizations. However, quality and regulatory affair employees in an organization need to make the designation jointly and document their reasoning in case of an audit.
If the FDA's Purolea letter is the new bar, what does “authorized human review” actually have to look like for something like a computationally generated protein structure figure — a checkbox, or a documented technical review?
The answer is closer to the latter. The Purolea letter didn’t just call out the absence of a human signature. It identified a fundamental issue in that there wasn’t a substantive human review documented. They used AI-generated documents without verifying compliance with CGMP requirements. Simply clicking a checkbox to provide a signature without actually doing a deep-dive review would likely yield the same scrutiny.
Developments across the industry are in flux. The FDA has issued a warning letter but not a rule. Annex 22 is still a draft. Regulations are still being developed and vendors are still trying to evolve with them.
However, workbenches, regulators, and vendors need to build a connective tissue between different technologies to ensure there is always human oversight behind any AI-generated decision.
This story was produced by Kivo and reviewed and distributed by Stacker.