On — Nov 13

Anthropic introduces Claude Meetings, which adds Claude to a team's video call as a visible participant. It builds docs, whiteboards, and other artifacts during the call, and it is rolling out in beta on Team and Enterprise plans.

Why it spiked. It puts Claude in direct competition with the many standalone meeting note-taker apps.

• @claudeai Announces Claude Meetings and the beta rollout on Team and Enterprise plans.
• @claudeai Says you can ask Claude questions out loud or in chat, and the team can see and correct what it builds as it takes shape.
• @robj3d3 Lists more than two dozen meeting note-taker products that the feature overlaps with.

Self-hosted Claude Managed Agents is in private beta. Companies can host the agent harness, sandboxes, and session history in their own cloud or on-premises, and choose where requests to Claude are sent.

Why it spiked. It lets enterprises give agents access to internal code and systems without that data leaving their infrastructure.

• @ClaudeDevs Announces the private beta for hosting the harness, sandboxes, and session history yourself.
• @ClaudeDevs Suggests uses like fixing broken builds or triaging alerts, and says access is for eligible orgs through their Anthropic account team.
• @katelyn_lesse Says Managed Agents is built as the API platform for powerful and secure agent systems, and companies can now host it themselves.

Vercel releases UI SDK, a library of more than 50 interface behaviors such as popups, selection, typeahead, and focus management. It ships components for React, Vue, Svelte, Solid, and Web Components, designed for agents to build with.

Why it spiked. It comes from shadcn, whose component library many developers already use, and it targets agents as the main builders of UI.

• @vercel Introduces the UI SDK.
• @shadcn Explains that agents now build and use interfaces, so the SDK packages behaviors like those behind a Select rather than finished components.
• @ctatedev Lists the 50+ behaviors and supported frameworks, with examples of what agents built using it.

At its Cyber at the Frontier livestream, OpenAI launches Daybreak Red Team for automated penetration testing and Daybreak Defense Factory for vulnerability handling and patching. A slide also shows an unreleased cyber model at 77.0% on ExploitGym and 79.1% on PatchBench; these are vendor-reported scores, and the model has no name, date, or price.

Why it spiked. Many viewers expected GPT-6.2 and instead got a benchmark slide showing a large jump over GPT-6 Cyber, which shipped about a week earlier.

• @gdb Announces Daybreak Defense Factory and Daybreak Red Team for putting frontier cyber capabilities to work for defenders.
• @thsottiaux Links the recording and says the new tools cover automated patching and automated red teaming.
• @chetaslua Posts the ExploitGym numbers from the slide: 77% for the unreleased model against 54.8% for Gemini 4 Carbon and 52.6% for GPT-6 Cyber.

Expo releases Expo Simulators, cloud iOS simulators and Android emulators built for coding agents. An agent runs its change on a cloud device and returns screenshots or a recording before it asks a human to review.

Why it spiked. Mobile developers have had no easy way for agents to test app changes on a device, and this gives agents that step in CI.

• @expo Introduces cloud simulators for agents that bring back a recording of each change.
• @expo Lists uses: PASS or FAIL with a recording on every PR, turning a TestFlight screenshot into a reproduction, and reviewing builds from a browser link.
• @nater02 Says agents can now start iOS or Android simulators in the cloud, use the app, and record screenshots and video as evidence.

World Labs opens access to its Atlas world model to everyone, with apps and API endpoints. Users start from an idea for a 3D world and get keyframes, video clips, or 3D Gaussian splats.

Why it spiked. Fei-Fei Li and investor Martin Casado promoted the launch, and a competing lab posted a side-by-side comparison the same day.

• @theworldlabs Opens Atlas to everyone with apps and API endpoints that output keyframes, video, or Gaussian splats.
• @martin_casado Calls it the largest step in 3D generation and reconstruction since the start of generative AI.
• @chuhang1122 Compares Atlas side by side with JST-2, which shipped on Oct 19 and generates each expansion in native 3D.

Continuous Intelligence, a London lab, releases ci-proto, its first time-series foundation model family. It reports first place on fev-bench ahead of models from Google, Amazon, and Mistral, a context window 64 times longer than others, and open weights for commercial use.

Why it spiked. A four-month-old lab tops a forecasting benchmark against much larger companies and releases the weights.

• @benjaminwalker Announces ci-proto with the fev-bench ranking, 64x context window, and open weights for commercial use.
• @DataChaz Points out that a small London lab released it as open weights and recaps the benchmark and context claims.