On September 25, 2026, Anthropic published two research papers on the same day. One documents a nine-loop scattering amplitude calculation in planar N=4 super-Yang-Mills theory. The other describes the autonomous discovery of a novel enzyme system with CRISPR-like repeat arrays in bacteriophage DNA. They are distinct problems in different fields, but the pattern is the same: Claude operating as a scientific co-researcher, not a chatbot.

The Physics Paper: Nine Loops in N=4 Super-Yang-Mills

The first paper, "Claude computes a nine-loop amplitude in N=4 super-Yang-Mills", is authored by Liam Fitzpatrick and Siddharth Mishra-Sharma at Anthropic with guest commentary by physicist Matt von Hippel. The result: Claude successfully computed the nine-loop MHV six-particle amplitude in planar N=4 super-Yang-Mills theory, extending work previously reported through eight loops.

Why this matters: scattering amplitudes in this theory are a proving ground for quantum field theory methods. Each additional loop represents an exponential increase in computational complexity. The nine-loop result was an open challenge issued to AI companies by the theoretical physics community. Claude reached it through two independent paths — a direct bootstrap calculation and an indirect method using the Steinmann relations — and both converged on the same answer. The result was then validated by human experts.

The paper is careful about what this demonstrates. It does not claim Claude "understands" physics in a human sense. It demonstrates that when a problem can be decomposed into formal reasoning steps and verified by computation, a language model with tool access can execute the reasoning chain at frontier level. The calculation involved symbolic manipulation, Feynman diagram evaluation, and algebraic simplification at a scale that would take human specialists months.

The Biology Paper: A CRISPR-Like System in Phages

The second paper, "Claude discovers a novel enzyme system with CRISPR-like repeats", comes from Anthropic's life sciences research program. The discovery: an enzyme system called array-associated reverse transcriptase (ART), found in bacteriophages, that combines a reverse transcriptase with a neighboring partner gene and a long array of evenly spaced DNA repeats — a pattern reminiscent of CRISPR arrays.

The process: approximately 950 Claude agents screened over 200,000 reverse transcriptases in 21 hours, flagging ART as a candidate. The repeat pattern resembles CRISPR arrays, suggesting the system may be programmable for DNA-related operations. Early experiments show the ART array expresses distinct short RNAs, similar to how CRISPR systems behave.

This is not a fully characterized gene-editing tool. It is a candidate system discovered through agentic search. The significance is the method: a fleet of agents with database access and reasoning capability conducted a systematic screen that would be impractical for human researchers to run manually. The discovery came from one of Anthropic's first dedicated research programs, suggesting this is the beginning of a sustained effort, not a one-off demo.

What Connects These Papers

Both papers share a structure: a problem that decomposes into formal reasoning steps, tool access (computational libraries for physics, genomic databases for biology), and a verifiable result. In physics, verification is mathematical consistency and expert review. In biology, verification is experimental follow-up — the paper notes early experimental validation of the RNA expression pattern.

Both also use agent fleets, not single prompts. The physics paper describes a multi-step reasoning chain with tool calls. The biology paper explicitly states ~950 agents ran in parallel. This is the same infrastructure that powers Claude Code, now applied to scientific research.

What This Means for Developers Building on the Stack

If you are building agents on the Anthropic stack, these papers are not just research showcases. They reveal the architecture Anthropic is investing in: multi-agent orchestration, tool use at scale, and verifiable reasoning chains. The same primitives — computer use, browser use, Files API, Agent Skills — are now available through the API and the Claude Agent SDK (released alongside Sonnet 4.5 in September).

Two practical takeaways:

  • Verification is the bottleneck, not generation. Both papers emphasize that the hard part is not getting the model to produce an answer — it is building the verification layer that makes the answer trustworthy. For physics that means mathematical cross-checks. For biology that means wet-lab follow-up. Any agent system you build for high-stakes domains needs an explicit verification strategy.
  • Agent fleets are the unit of work, not single agents. The biology paper's 950-agent screen is the clearest signal yet that Anthropic views parallel agent orchestration as the default for complex tasks. The Claude Agent SDK exposes this pattern. If your architecture still assumes one agent per task, you are designing for the previous generation.

The Credentialed Channel Pattern

Anthropic's framing is notable: they are not loosening safeguards globally. They are building a credentialed channel — capability-plus-gate — where advanced research capabilities are available through a verified pathway. This converts safety posture into an enterprise and regulated-market distribution mechanism. For developers, it means the most capable agent features will come with access controls, audit trails, and compliance tooling built in.

The Bottom Line

These two papers, published on the same day, are a signal. Anthropic is demonstrating that its models can contribute to frontier scientific research in both theoretical physics and experimental biology. The common thread is not "AI does science" — it is that problems decomposable into verifiable reasoning chains with tool access are now tractable for language-model-driven agent systems.

For builders on this stack, the question is not whether to adopt agent orchestration. It is whether your verification layer can keep up with what your agents can generate.

If this was useful, you can support my open-source work on Ko-fi or check out my services.