Claude's biology lab finds a new enzyme system

Anthropic gave Claude a vast DNA search and got back an enzyme system that scientists had not described before.

Today's issue covers AI-led biology, Amazon's always-on seller agent and Meta's new connector market.

Anthropic has built a molecular biology lab to test whether a general AI system can find promising biology that human scientists would otherwise miss.

Its first published result is a previously uncharacterised enzyme system associated with repeating DNA sequences. The pattern resembles systems such as CRISPR, though Anthropic does not yet know what the new system does.

The search was large:

The result is narrower than an autonomous scientific breakthrough. Claude found a candidate worth testing. The wet lab still had to show that the pieces belonged together, and the biological function remains unresolved.

AI can search more material than a small team could cover, while experiments decide whether the output survives contact with the real world.

Founders context: If you are buying AI-assisted discovery work, price the search and the proof separately. A promising candidate is an input to validation, not the finished result.

Amazon's Seller Assistant is moving from answering questions to running repeat jobs while the seller is offline.

Sellers can describe a workflow in plain language, choose its guardrails and decide whether it should recommend an action or carry it out after approval. Every action gets an audit trail.

The first jobs are ordinary but frequent:

Amazon says the assistant now reaches more than 90% of its selling partners, has hundreds of thousands of active users and sees recommendations accepted more than 90% of the time.

The new memory carries pricing patterns, inventory cycles and growth goals between Seller Central, Amazon Quick and a Claude beta plugin. Amazon says seller data remains inside its infrastructure.

This turns the product from a chat window into a monitored operating loop. The controls matter as much as the model because the assistant can now touch prices, stock and customer-facing listings.

Founders context: Start a workflow agent on one repeat job with an approval step and a complete log. Remove the approval only after its mistakes are boring, visible and reversible.

Meta has opened Muse, its personal AI agent, to outside services through connectors. A connector lets Muse call a product when a user asks for help, without sending that person to a separate app first.

Greg Isenberg, who hosts the Startup Ideas Podcast, used Thursday's episode to work through four possible services:

The common shape is small. Each connector does one job, exposes a clean action to Muse and passes Meta's review.

A first version can be built with a coding agent before the team invests in a full product.

Distribution is the harder part. Meta's directory may surface a connector, but Isenberg argues that creators and existing customer relationships will still matter. A listing does not guarantee demand.

The commercial tension is familiar from earlier app stores. Connectors put a service close to user intent, while Meta controls approval, discovery and the interface. The opportunity arrives with platform dependence already attached.

🛰 Google prepares to test AI hardware in orbit. Project Suncatcher will launch a prototype satellite to test radiation, vibration and cooling before Google considers larger orbital clusters.

💰 Akamai signs an $11.6bn compute agreement with Anthropic. The multi-year CPU deal includes a warrant for up to 5% of Akamai, widening the infrastructure race beyond GPUs.

🏥 AI scribes are linked to $942m in extra health costs. The Blue Cross analysis says richer documentation can increase coding and reimbursement, so saved clinician time does not automatically lower the bill.

🎧 AI Daily Brief tracks agents moving into devices. Nathaniel Whittemore's 24 September episode follows consumer agents as they move into glasses, cars and routine administration.

📏 Anthropic proposes measures for AI development speed. The research group wants labs to track how quickly systems improve and how much AI contributes to building the next generation.