Latest briefing
June 30, 2026 · 5 lead stories (site) · 7 stories (base)
5 lead stories · 4 to watch. The June 30, 2026 briefing focuses on the invisible infrastructure that turns AI agents into genuinely deployable enterprise tools: identity standards (x401), governance (Josys), no-code platforms (Pentagon), inference infrastructure (Baseten) and price diplomacy (California / Anthropic). Beat: pure agentic — the software stack that converts an agent demo into a production tool.
🔥 Top story
01
A new open standard wants to know "who is behind this AI agent" when it acts on your behalf
Imagine your AI assistant placing an order, transferring €200, or signing a document for you. Today, the website or bank on the other end has no standard way to verify that a real person actually authorized that action. That is exactly the gap Proof — a company that already secures $640 billion in transactions per year — decided to fill with x401. On June 25 it published this open protocol: a little like the badge of a corporate building, internet-style, that cryptographically proves a human actually authorized this agent, within a precise scope (you authorize it to spend up to €500 per day, but no more). The project is backed by Circle (the same company behind agentic payments), OpenAI, Google and Okta, and submitted to the FIDO Alliance standards group. If x401 becomes the standard, it is the identity layer the agentic internet was missing — the one that makes it possible to trust an agent acting in your name.
02
The Pentagon puts a no-code AI agent builder in the hands of its 3 million personnel
When the Pentagon allows 3 million people — soldiers, analysts, accountants, administrative staff — to build their own AI agent without writing a single line of code, "no-code agent builder" becomes a de facto standard. That is what happened at the end of June with "Agent Designer", a tool integrated into GenAI.mil, the official platform of the U.S. Department of War (the new name for the Department of Defense). The tool runs on Google's Gemini models and lets anyone build an assistant that executes a multi-step task, draws on several data sources, and can be shared with a team. Use cases already tested: automatic after-action reports after an operation, synthesis of classified images into memos, mini financial-analysis apps. For the general public, this is proof that AI agents are no longer reserved for engineers — they are becoming a routine business tool, like a spreadsheet or an email, open to everyone.
03
California signs with Anthropic: Claude at -50% for its 230,000 civil servants
Governor Gavin Newsom announced on June 30 an agreement with Anthropic: Claude becomes the first productive AI assistant available to all California state agencies, with 50% off the catalogue price, free training and direct support from Anthropic engineers. California cities and counties (local authorities) get the same discount, opening the deal to about 230,000 public servants in total. The state already runs Claude on Poppy (its public-facing AI assistant), DMV services (the U.S. equivalent of French « préfectures »), Medicaid and the cybersecurity of state code. The deal lands in the middle of the tension between Washington and Anthropic — the White House restricted use of the latest Claude Fable 5 and Mythos 5 models on June 12. Newsom says this is not a snub to Trump, but the signal is clear: U.S. states can now negotiate their own AI pricing, independently of federal policy.
04
Baseten raises $1.5 billion to become the "AWS of AI inference"
If Amazon rents computing power to anyone wanting to run their websites, Baseten wants to do exactly the same thing for AI agents: rent them inference capacity (the ability to run an AI model, as opposed to training it). On June 22, the platform closed a $1.5 billion funding round, valuing the company at $13 billion — one of the largest rounds in AI infrastructure history. Behind the curtain, Baseten already runs Cursor (the coding agent), Clay, Lovable, Mercor, OpenEvidence — more than a billion inference calls a day, ×20 in one year. The signal for companies: investors' money is leaving model training to flow toward the infrastructure that serves them in production. Concretely, this means the cost of using AI agents is becoming a commodity — like electricity, on-demand and at the meter.
05
Josys launches a dashboard to know exactly which AI agents run inside your company
In most large companies today, nobody really knows how many AI agents are running, who deployed them, what data they can access, or whether the IT team approved them. It is a bit like having building access badges without knowing who wears them. On June 29 Josys — a Japanese platform specialized in digital identity governance — launched a new dashboard that answers exactly that question. The tool automatically discovers all AI agents present (Microsoft Copilot, Claude agents, in-house custom agents), classifies them by owner, flags "shadow agents" (created without approval) and those with too many privileges. When the "AI agent governance" market is being structured like this, it is a sign that agents are leaving the experimental phase to become a standalone IT subject — like users, servers or applications.
📡 To watch
Stark Defence: $540M for AI-piloted drones — the far end of the "autonomous agent" spectrum is now funded
On June 23, Berlin-based autonomous loitering drone maker Stark Defence closed a $540 million round led by Sequoia and Founders Fund (Peter Thiel), valuing the company at $3.65 billion. Its "Virtus" drones have a range of more than 130 km, and a single operator can fly an entire swarm through "Minerva" software. Beyond the military, this is evidence that investors fund autonomy at scale when the market demands it. For civilian companies, the signal is that the whole agent industry will structure itself around this precedent: "semi-autonomous", "supervised", "advise-only".
Peregrine: "privacy-first" becomes the regulatory standard for public-sector AI
Peregrine — the platform already serving 400+ North American public agencies (Super Bowl, Grammy, World Cup) without creating new personal data — closed on June 22 a $250M round at a $6.8B valuation. The public-sector AI market is forecast at $109B by 2035. The model — exploiting existing data under explicit permissions rather than creating new data — is becoming the GDPR-compatible standard in Europe and the NOR-compatible standard in Norway.
California / Anthropic: "discount diplomacy" could spread to OpenAI, Google and Mistral
Newsom indicated he wants to sign similar deals with other AI providers. If OpenAI, Google or Mistral follow suit, California becomes a frontier-AI market independent of federal policy, with predictable negotiated pricing. This precedent can be exported: European regions, Canadian provinces, Brazilian states.
GLM-5.2 (Z.ai): Chinese open-source weights still pending after the California deal
GLM-5.2, published on June 13 by Z.ai with very strong benchmark results (62.1% on SWE-Bench Pro), still has to publish its open-source weights. With California locking down its Anthropic access, pressure mounts for the Chinese open-weight community to publish weights — a digital sovereignty dynamic that could accelerate in the coming weeks.
📊 Trend
June 30, 2026 marks a silent but deep shift: for the first time, agent identity standards (Proof x401), enterprise governance tools (Josys) and no-code build platforms (Pentagon Agent Designer) are all maturing at once, as if the whole ecosystem had an appointment to move from "agent as a gadget" to "agent as critical infrastructure". At the same time, investor money is leaving model training and flowing toward inference infrastructure (Baseten) and military agentics (Stark Defence), while U.S. states — California first — start setting their own rules of the game. The AI agent market increasingly resembles the cloud computing market fifteen years ago: usage explodes, prices fall, and the question is no longer "will it work" but "who will use it first".