Perplexity releases Search SDK for agentic harnesses

#1 𝕏

Aravind Srinivas says Perplexity released its Search SDK for use inside agentic harnesses, describing it as the technology that makes Perplexity Computer a best-in-class product for wide and deep research.

#2 𝕏

Cursor announced that its acquisition by SpaceX has officially closed and that it will join the SpaceXAI team. The company said it will help improve Grok Build, Grok Bot, Grok API, Cursor, and other offerings.

#3 𝕏

Josh Woodward announced that 3.7 Flash is available in the @GeminiApp.

#4 𝕏

A post announced that Qwen3.8-2.4T-A95B is live on SiliconFlow, describing it as offering “Max-level intelligence” on “Day 0.”

#5 𝕏

NVIDIA recapped NVIDIA NeMo Switchyard, described as a new open-source library for routing agent tasks across models. The post recommends frontier models for complex reasoning and planning and NVIDIA Nemotron Lightning for high-volume, specialized execution.

#6 𝕏

Josh Woodward recapped Google’s release of Credentio as open source, a new C++ library for C2PA Content Credentials intended to help developers build better provenance tools.

#7 𝕏

Hugging Face shared “The State of Open Models, Summer 2026,” reporting that frontier models are getting larger while small models still dominate real-world usage. Qwen leads local inference ahead of Gemma, and AI agents are becoming a major force on the Hub.

#8 ▶️

AgentMail Founder Adi Singh on Email for Agents

SyntaxGTM

#9 𝕏

Santiago shared an unnamed library that can mock LLMs, MCP tools, agents, databases, and other components. It can be installed or run as a Docker container, enabling tests without hitting real components.

#10 𝕏

Ali Ghodsi shared how Smart Routing works on the author’s AI Gateway, describing it as a “super simple idea” that lowers costs by about 30% without sacrificing quality.

#11 𝕏

Thariq said turning an inference API into an external API is surprisingly difficult, as implementations can drop thinking blocks or disrupt cache behavior—especially when serving across different providers.

#12 𝕏

Ali Ghodsi commented that Genie uses org-chart and popularity signals to select the “right” answer, surfaces uncertainty, and remembers user answers across future sessions to improve over time.

#13 𝕏

Harrison Chase shared new LangSmith documentation covering traces, threads, and trajectories—a new concept—and said observability data can support memory and learning, making a clear mental model especially helpful.

#14 𝕏

Anthropic announced that its second Risk Report is now available, detailing risks from its systems and its preparedness to address them as part of its Responsible Scaling Policy.

#15 𝕏

Anthropic wrote an FAQ explaining that it is implementing watermarking to comply with the EU AI Act. The company says the method will not affect Claude’s output quality or cost, alter text or add hidden characters, be perceptible to readers, or enable tracing to a specific person, organization, or chat.

#16 𝕏

Qwen3.8-27B is available on Ollama, and the author invited users to share what they build with the model.

#17 𝕏

Google AI recapped new AI integrations for the Pixel 11 series, Pixel Watch 5, and Pixel Tag, alongside Gemini 3.7 Flash’s availability in the Gemini API and rollout to paid users. It also highlighted Google DeepMind’s open-source WeatherNext 2, which can provide an extra day of accuracy in cyclone forecasting, and the rollout of the upgraded Gemini Notebook experience to all Pro users and notebook copying to all users.

#18 𝕏

Garry Tan compared using GStack before and after Fable 5, saying that for many one-way-door questions returned in Claude Code, users can simply choose “Take all recommendations” and be happy.

#19 𝕏

Tal Raviv says OpenAI released “Computer History,” describing it as OpenAI’s version of Familiar. He highlights Familiar as an offline, local, open-source, free, and model-agnostic option for screen history.

#20 𝕏

Andrew Ng shared a link described as a map of the most important skills in AI Engineering.