Three of today's developments concern the same underlying question: which real-world systems are AI agents now authorised to operate in. A fourth offers a practical instrument for teams building the agents themselves.

Claude identifies a novel enzyme system in 21 hours of autonomous research

Anthropic's life sciences research group, formed in spring 2026, ran an autonomous campaign in which 950 Claude agents analysed a database of roughly 1.9 billion protein clusters. The campaign lasted 21.5 hours, consumed 210 million tokens, and produced one verified candidate: a previously uncharacterised enzyme system found in bacteriophages, which Anthropic calls array-associated reverse transcriptases (ART).

The ART system has three components: a reverse transcriptase enzyme, a partner gene, and a long array of evenly spaced DNA repeat sequences whose layout resembles a CRISPR array. The configuration suggests the system performs programmable DNA operations, though its exact biological function is still under investigation. All wet-lab validation work is performed by human scientists in Anthropic's Bay Area BSL-1/2 facility.

Feng Zhang, a CRISPR pioneer at MIT's Broad Institute, described it as "an exciting example of how AI agents can contribute to biological discovery." Anthropic says a conventional screening campaign of this scope typically takes weeks to months. The operative claim for operators: the scanning, filtering, and candidate-ranking stages of a scientific discovery workflow can now be delegated to a large fleet of agents at a cost measured in compute, not research hours.

Amazon opens Seller Central to outside AI agents via a Claude plugin

At its Accelerate 2026 seller conference on 23 September, Amazon announced the next evolution of Seller Assistant: a new Selling Partner plugin that gives Claude and Amazon's own Quick assistant programmatic access to Seller Central data. Sellers can now manage inventory, pricing, listings and analytics through a Claude session without opening the Seller Central dashboard.

The assistant, built on Amazon Bedrock with Anthropic's Claude models, carries persistent memory of each seller's pricing patterns, inventory cycles and growth goals. That memory travels between Seller Central, Claude and Quick instead of resetting between sessions. Seller data stays within Amazon's infrastructure and is not shared externally. The plugin is in beta for US sellers, with international expansion to follow. Amazon is offering every primary account holder globally a free 12-month Quick Plus subscription through 31 December 2026.

The adoption baseline gives context: Seller Assistant is already live for more than 90 per cent of Amazon's selling partners worldwide, with users accepting its recommendations more than 90 per cent of the time. The plugin step is materially different from a chat feature. It opens a high-traffic commercial backend to an external agent framework for the first time. The precedent matters more than the feature.

Tekever closes $580 million Series D at a $6.4 billion valuation

The Portuguese-British AI drone maker Tekever announced the first close of a $580 million Series D on 23 September, led by UC Investments (the University of California's investment arm) and Baillie Gifford, with Merlyn Advisors joining as a new investor. The round marks UC Investments' first direct investment in Europe.

Tekever's drones have accumulated over 50,000 operational flight hours in Ukraine since Russia's full-scale invasion in 2022. One week before the announcement, the company signed a contract worth up to £400 million to replace the British Army's Watchkeeper surveillance drone fleet. Proceeds will fund geographic expansion, industrial capacity and acquisitions.

  • Baillie Gifford's presence signals long-duration institutional capital, not typical venture positioning.
  • UC Investments crossing into European defence tech for the first time reflects a structural shift in where US university endowments are deploying capital.
  • The valuation implies the market is pricing operational track record in live conflict as a durable competitive advantage, not a temporary premium.

Google open-sources EnvHarness for adaptive agent training

Google Cloud AI Research released EnvHarness under Apache 2.0 on 20 September. The framework places a programmable wrapper around existing training environments, adjusting start positions, visible observations, allowed actions and episode length without modifying the underlying verifier. A companion tool, EnvRigger, observes an agent's execution trajectories and synthesises difficulty adjustments that target its specific weaknesses.

Across five benchmarks spanning software engineering, web navigation, office automation and embodied tasks, agents trained with EnvHarness improved by up to 9 points on held-out evaluations. On SWE-bench Verified, the score moves from 52.13 to 54.79. The framework exposes three mechanisms: Stage (where an episode starts), Contract (filters on actions or observations), and Chain (linking tasks into longer episodes).

For teams building coding or enterprise agents: EnvHarness is environment-agnostic, requires no changes to the verifier, and the benchmark delta is meaningful on a leaderboard where frontier models compete within a few percentage points. The open licence removes a barrier that has historically kept adaptive-curriculum methods inside large labs.

The through-line across today's brief is one of access. Agents are being handed control over systems that were previously gated behind human intermediaries: biology databases, commerce backends, surveillance platforms, training curricula. The pace at which organisations open those gates will determine whose AI compounds fastest. An operator who has not mapped which of their own systems could be opened to agents is already behind the planning curve.