join.dev
AI news · 10 stories
Moonshot releases Kimi K3.1 Pro with million-token context · Claude scales formal verification across 123 open-source libraries · Agentic AI Foundation adds four open agent projects
Moonshot released Kimi K3.1 Pro and a faster Highspeed variant across its apps and API, with a one-million-token context window. Full weights are promised by November 4, extending future self-hosting options for coding and knowledge work.
Read at Moonshot AI
Anthropic and Theorem used Claude to produce checked proofs for 239 test entry points across 123 open-source libraries. The work demonstrates broader automated verification, but guarantees apply to specified RISC-V binaries and assumptions, not every bug or deployment.
Read at Anthropic
The Agentic AI Foundation accepted Agent Skills, Open Responses, OpenEnv and OpenSandbox on October 22. The contributions extend neutral governance across reusable capabilities, interoperable interfaces, training environments and agent execution.
Read at Linux Foundation
Google announced Agent Compute Engine, giving agents hardware-isolated microVMs with persistent storage and identity controls; customer preview is planned in the coming months. It also introduced the open-source Agent Substrate runtime for Kubernetes, addressing long-running agent execution.
Read at Google Cloud
1Password announced 1Password for Agents to let supported personal agents use approved logins without seeing or storing the credentials. The platform keeps secrets in the vault while brokering access, reducing the need to paste passwords into agent conversations.
Read at 1Password
Sakana AI announced Aquarium on October 23, connecting organizational data, rules and knowledge through an ontology. Managed workflows combine agent actions with permissions, approvals and audit records, supporting reusable enterprise automation.
Read at Sakana AICopilot usage reports now identify custom skills, agents, MCP servers, commands and plugins by configured name, with up to 25 entries per record. The change helps administrators track adoption, although client coverage varies and earlier reports remain unchanged.
Read at GitHub
GitHub consolidated notifications and Copilot into one Teams app, letting a mention turn an issue notification into an investigated fix and pull request. Teams still review, test and merge changes; the legacy Copilot app is scheduled to retire November 16.
Read at GitHub
An October 22 preprint shows that deleting a memory entry can leave secrets recoverable from derived agent state. Removing dependent information eliminated observed recovery in tested settings, highlighting why deletion needs behavioral checks beyond storage operations.
Read at arXiv
An October 22 preprint compared six open models across five domains and five prompt wordings, finding that tools improved some accuracy results while destabilizing many pairwise rankings. The authors recommend reporting tool-call frequency and ranking stability alongside accuracy.
Read at arXiv