Daily briefing
July 1, 2026 · 4 stories (site) · 8 stories (base)
🔥 Top story
01
Nvidia releases a complete free toolkit so anyone can build their own AI agents
Imagine you want to build an assistant that can search for information, place orders, fill in forms on your behalf — in short, an "AI agent." Today, you either pay a lot for a ready-made solution, or assemble dozens of mismatched parts yourself. Nvidia just published the entire toolkit at once, for free: a framework to orchestrate multiple agents together, shared memory between them, a secure space to run them safely, and a language model with 550 billion parameters that's said to be five times faster than the previous generation. It's a bit like a car manufacturer publishing the engine, the chassis, the dashboard and the gearbox all under an open license. The first to use it are names you already know: Perplexity, Palantir, ServiceNow, CrowdStrike. For a company like Info-Sys that already uses Claude Code every day, this means there's now a credible fallback option if the American editor becomes inaccessible — for example because of a new export restriction like the one on June 12.
02
Anthropic lets large companies keep Claude Code locked inside their own cloud
Many IT departments refuse to use an AI assistant because it sends data outside. Anthropic just solved this problem for large enterprises: you can now run Claude Code behind the walls of the company's own cloud — whether Amazon's, Google's or Microsoft's — while keeping the same connection to your enterprise account. In practice, it's as if you could rent the muscles of a world-class athlete without ever letting them leave your gym. The system handles employee authentication, team-level rules, per-user billing, and sends activity logs in the industry-standard format. For a tech team like Info-Sys that already uses Claude Code Pro every day, this is the gateway that was missing to move from experimental use to a real production deployment at clients with strong security constraints.
03
Anthropic launches a "Claude for researchers" that handles 60 lab tools and double-checks its own sources
Doing scientific research means jumping from one tool to the next all day: PubMed for papers, Jupyter for calculations, the terminal to launch simulations, UniProt for proteins, and so on. Claude Science pulls all of that into a single assistant that knows which tool to use for which task, and that has its answers reviewed by a second agent whose job is to verify that citations are accurate and calculations are correct. For a biology lab, it's the equivalent of a tireless research assistant who never gets the index wrong and never invents a reference. Anthropic isn't stopping at life sciences: the model is designed to be extended to other domains — finance, law, engineering. For anyone building products, this is proof that the same agent engine can be specialized by industry and become a product in its own right.
04
A Japanese startup offers an assistant that decides for itself who to delegate the work to: ChatGPT, Claude, Gemini, or itself
Today, choosing an AI assistant means picking a vendor — and sticking with it. Japanese startup Sakana takes a different approach: one single assistant that decides for itself, for each question, whether to answer directly or hand off the work to another model — whether GPT, Claude, Gemini, or one of its own. It's like having a project manager who knows each team member's specialty and who knows who to turn to depending on the nature of the problem. The pitch is as much political as technical: Sakana claims performance on par with the best American models without depending on any single one of them — an explicit answer to the export restrictions that hit some models in June. For anyone following the topic closely, this confirms that the monolithic model is giving way to the orchestrator.
📡 To watch
Claude Sonnet 5: 5 to 7 times cheaper than before, free for individuals — the agent economy tilts
Anthropic launched Claude Sonnet 5, billed as "the most agentic" of the Sonnet range, and placed as the default model for both free and paid consumer plans. The entry price is five to seven times lower than the top-of-the-line Opus for very close performance on common agent tasks. For a stack that consumes a lot of tokens, this is a major shift in price-for-performance: until the end of August, the introductory rate is $2 per million tokens read and $10 per million tokens written. If the benchmarks hold in real conditions, it's a signal that switching to Sonnet 5 for non-critical workloads becomes a no-brainer.
Will Anthropic extend Claude Science beyond biology?
Claude Science is currently only available in beta for life sciences. If Anthropic extends it to finance, law or engineering, it would signal a move into a range of "Claude Cowork specialized by industry." Worth watching closely, since this is a product model replicable by any agent editor.
The coalition around Nvidia is growing — who controls the open-source agent stack by end of 2026?
Seven initial partners have already joined the NemoClaw coalition (Perplexity, Palantir, ServiceNow, CrowdStrike, Glean, Kilo Code, LangChain Deep Agents). If the coalition reaches twenty members by year-end, Nvidia becomes de facto the reference editor for the open-source agent stack on the Western side, with a weight comparable to Linux on servers.
Will the "trust box" co-built by Nvidia, Microsoft, Canonical and Red Hat become the standard?
Running an agent in a space fully isolated from the rest of the system is the bedrock of security. OpenShell Secure Runtime, announced jointly by Nvidia, Microsoft, Canonical and Red Hat, targets exactly that role. If it's adopted by actors beyond the early users, it could become the de facto standard for running agents in production — on par with Google's and Amazon's technologies on their respective clouds.
📊 Trend
July 1, 2026 confirms that the agent stack is structuring itself at speed. Four layers are emerging simultaneously: models (Nvidia's Nemotron 3 Ultra and Anthropic's Claude Sonnet 5 are 5 to 7 times cheaper), orchestration tools (Sakana Fugu plays the "one for all" card, NemoClaw publishes the full stack), trust boxes (OpenShell Secure Runtime becomes a candidate for the de facto standard), and verticalized products (Claude Science proves that one engine can become a per-industry product). Consequence for anyone building with AI: dependence on a single editor has never been riskier — nor easier to reduce.