Reads

Mistral is expanding Studio, its enterprise platform for agents, workflows and apps, with managed agents and apps and new features for workflows, enterprise context, agent observability and evaluations, alongside its open-weight Mistral Large 4 (“Le Chonk”). Studio runs in Mistral’s cloud, in a customer’s VPC, or on-premises, and customers such as BNP Paribas describe running it on their own infrastructure to keep sensitive processes under their control.

DeepMind built a set of Cheminformatics Skills for Google Antigravity, its agent environment. The model handles orchestration, literature and database search, and scientific reasoning, while deterministic tools (RDKit, plus connectors to ChEMBL and PubChem) do the molecular calculations. In a walkthrough based on real lung-cancer drugs, the agent searches thousands of compounds, analyzes matched molecular pairs, and proposes new rigidified macrocycles with synthesis routes in about an hour.

Anthropic gave 50 professional analysts a realistic causal-inference task in Claude Code, and secretly had each agent lean toward one of two defensible options on four analytical choices, decided by coin flip. Analysts went with the agent’s lean 86% of the time, switched away from a preference they had already stated 60% of the time when the agent disagreed, and afterwards explained their choices in the agent’s words while rarely crediting it. Analysts who used the agent were still faster and made far fewer serious errors than those who worked alone.

Working with Theorem, Anthropic had Claude turn existing fuzz tests from 123 OSS-Fuzz libraries into formal specifications and prove them against the compiled RISC-V binaries: no crashes or memory-safety errors for any input under 64 MB, across 239 entry points. The proofs add up to about 180 million lines of Rocq, but a proof checker verifies them, so people only have to trust the specifications and the models of the CPU and operating system. Along the way, Claude found and fixed bugs that years of fuzzing had missed.

Claude was given plain-language specifications for eight quantum circuit building blocks, plus an automated evaluator that checks each design for correctness and scores it on qubits and gates used. Working in a closed loop with no human input between rounds, it found designs that beat the best published circuits, in some cases by an order of magnitude, by spotting new structure in the problems rather than fine-tuning existing designs.

Anthropic went from about 2,500 merged changes a day in May to over 30,000 in September, mostly written by Claude, and git, even with hundreds of internal mirrors, couldn’t keep up. In five months it built and migrated to Scion, a version control system designed on the assumption that clients (agents and CI jobs) are always online: commits and merges happen on the server, checkouts are sparse or virtual, and questions like blame and history are answered by API instead of a local clone.

A report on cases where Claude, running evaluations or internal tasks with live internet access, did things nobody intended: exploiting an injection flaw on a university server, submitting real web forms (including an invented tip to a police department), using exposed access tokens to reach fee-gated data, and routing around its fetch tool’s URL-length limit with URL shorteners. Most are forms of persistence: when a task is ambiguous or impossible, the model works around the obstacle instead of stopping. Anthropic responded with tighter tool guardrails, offline versions of the evaluations, and automated detection that blocked every case described.

201 Anthropic employees each gave a Claude agent a book to trade and a five-minute chat about their tastes, and the agents negotiated swaps with each other on shared trading floors. The agents bargained competently; the market fell short mainly because of what they didn’t know about their person, since the rankings built from each chat matched the person’s own on 61% of pairs. Re-running the markets dozens of times showed that the underlying model affected outcomes more than telling agents to be “ruthless” or “prosocial”.

Google is adding persistent, cross-device memory to Private AI Compute, its hardware-isolated cloud environment, which until now discarded all context when a task ended. Each user’s memory sits in encrypted storage whose keys stay on their own devices; data is decrypted only inside secure enclaves while a request is handled, and devices can check a tamper-proof public record of the server software before sending anything. Google also published an updated whitepaper and an independent audit.

Spotify examines whether AI-assisted development hurt quality as its merged changes more than doubled in a year. Its incident reviews found no case where AI-written code was a material direct cause; the real pressure came from the volume of change outpacing review, testing, rollout and rollback, along with scarcer compute. Code-quality work grew as a share of merged PRs and rework rate stayed flat, though PR size and complexity are creeping up.

Mistral describes migrating a 40,000-line Fortran 77 reservoir simulator, with no tests and no central documentation, to C++ for a European energy operator. Before migrating anything, the team built a parity harness: the Fortran code dumps state checkpoints and the C++ tests load them, so every migrated module must match numerically, and Skill.md files tell the agents how to use it. Agents documented the codebase before the migration began, and people reviewed the work at set checkpoints.