Anthropic opens its agent features and publishes protein research while its text watermark faces scrutiny
Anthropic advanced the Claude platform with general availability for agent tools and end-to-end protein design experiments, while the community rushes to build tools that strip its new invisible text watermark.
Artificial Intelligence··Midday
Protein pipeline automation
Anthropic published two experiments in which Claude independently coordinated open-source protein-design tools to construct minibinders, small proteins designed to latch onto specific targets. Out of 1,320 designs, 354 successfully bound to their target in laboratory measurements, translating to a nearly 27 per cent success rate across the set. The agent's top-ranked outputs performed even better, reaching a 49 per cent success rate. The company highlighted these figures against typical industry baselines, noting that a multi-target campaign run by the model carries estimated computing costs of 50,000 dollars.[1]
Platform features exit beta
Alongside the scientific demonstrations, the company removed the beta designation from its Files API and Agent Skills endpoints. These capabilities are now generally available on the Claude API alongside the Claude Enterprise user management tools, allowing standard requests without beta headers. For administrative control, the update allows teams using Claude Managed Agents to strictly regulate the domains the agent's search and page-fetching tools can access, relying on allowed and blocked lists to enforce boundaries.[2]
Watermark removal tools gain traction
As Anthropic's platform capabilities expand, its compliance mechanisms are meeting resistance. Following the company's decision to mark all Claude text outputs to meet requirements of the European Union's AI Act, online searches for watermark removal tools jumped 60 per cent in a single week. Developers are actively building scrubbing applications to counter the marking technique, driving an open-source project named Watermarks Remover past 15,000 stars on GitHub, even as the company insists the invisible marker makes no claims about actual ownership.[3]
Related columns
For more information on this topic, you can read the related columns.
