Latest briefing
July 3, 2026 · 5 items (site) · 5 items (base)
On July 3, 2026, AI agents shift from prototype to infrastructure: 100,000 agents deployed at the Pentagon, a $6.3 billion SpaceX-Reflection mega-deal, an open-source stack to audit agent risks — the agentic stack is coming of age.
🔥 Top story
01
The Pentagon puts 100,000+ AI agents on a single network to help commanders decide in seconds
When a military commander has to pick a target within minutes, they pull together analysts, satellites, drones and field reports. Tomorrow, they will ask a network of AI agents. On June 26, the U.S. Department of War announced "Agent Network", a system of operational agents that turns raw intelligence into decision options, with a human approving at the end. It is the second major project of the Pentagon's AI acceleration strategy. On July 1, the Pentagon's AI lead confirmed agents now automate the paperwork to validate a new piece of software — up to two years of process collapsed to a few hours. Three and a half million people already use these tools at the department. For the general public, this proves we have entered the era of "agents as infrastructure": no longer a tool you consult, but a team of digital co-workers preparing the work while you sleep.
02
Exabeam releases the first open-source risk-rating framework for AI agents, aligned with OWASP standards
Today, when you entrust a task to an AI assistant — writing an email, moving money, booking a trip — no one checks whether it has the right to do so before it acts. Exabeam, the global specialist in cyberattack detection, published on July 1 two pieces of free software that change the game: Observra reads what Claude Code, Codex, Gemini and Copilot agents do and sorts their actions into a common format; Praxen checks each authorization before the agent executes. Think of it as an automatic anti-fraud filter built into all your assistants. Fifty new monitoring scenarios join the existing forty, and every behaviour is now scored against the top ten agent-AI risks listed by OWASP. For any company that wants to ship an agentic product, this confirms security is becoming a mandatory layer — and a market to seize for whoever audits it.
03
SpaceX signs a $6.3 billion deal with Reflection AI to become the third superpower of AI compute
SpaceX, the company that launches Falcon rockets, is doing something strange today: renting out its own AI servers. On June 22, the company signed a $150-million-per-month contract — $6.3 billion over three years — with Reflection AI, an American lab specializing in "open" AI models. The latest-generation Nvidia GPUs are installed in its Colossus 2 data center in Memphis. This is already the third mega-deal of its kind, after Anthropic and Google. In total, SpaceX pulls in $27 billion a year in compute revenue, without ever calling itself an AI company. For the general public, this gives a new mental model: a data center can become a financial asset in its own right, like a downtown office tower. And American "open" models, so far overshadowed by Chinese models, suddenly have the war chest they needed.
04
GitHub finally lets you pick an open Chinese AI model as your coding assistant — a world first
When you use GitHub Copilot to write code, you choose today between closed American models. Since July 1, paying subscribers can also select Kimi K2.7 Code, an open model developed by the Chinese startup Moonshot. One trillion parameters in total, 32 billion active per request, a 256,000-character context window — and the ability to spin up to 300 sub-agents to slice a complex task. The model is hosted on Microsoft Azure, which is a discreet diplomatic gesture amid Washington–Beijing tensions. For European or Canadian developers who want to diversify their stack without depending only on Claude or GPT, this is a credible option. For the market, it proves the "Chinese vs American models" frontier is starting to blur as soon as a neutral cloud hosts both.
05
Ory releases identity plugins that drop straight into Claude Code, Codex and Gemini — authentication goes invisible
Today, an AI agent that accesses your bank data has to be authenticated like any employee. Problem: nobody has yet added the "robot" field to identity forms. Ory, which already manages 2.5 billion digital accounts worldwide, released on June 10 Agent DX, a collection of free plugins that automatically inject authentication, permissions and traceability into the most-used code assistants. Installation is a single command line; the rest is generated by the AI itself. For a startup building an agentic product, it is like receiving a preconfigured access badge on day one instead of running to security afterwards. For the industry, it is a signal: digital identity for agents becomes a standard of its own, and whoever imposes it will set the rules for the next decade.
📡 To watch
The chip embargo pushes China to innovate through software: the inference race is on
DeepSeek released on June 27 DSpark, a free software that speeds up its models by 60 to 85 percent. Combined with the release of MiMo-V2.5 (over 1,000 tokens per second), this confirms that China now bets on code efficiency to make up for restricted access to high-end GPUs. To watch: adoption in Europe via vLLM and SGLang, first independent benchmarks, the response from Western players.
The AI agent evaluation market is consolidating — who will be the "Moody's of agents" by 2027?
Exabeam Observra/Praxen is not alone: LangSmith, Arize, Langfuse, Helicone, WhyLabs and Fiddler are all positioning on the same turf. The market will likely concentrate around two or three leaders. To watch: publisher acquisitions, official partnerships with OpenAI and Anthropic, competing open-source projects (OpenAI Evals, HuggingFace LightEval).
Microsoft hosts Kimi on Azure: a political signal as much as a commercial contract
Amid Sino-American tension, seeing a Moonshot model run on Microsoft servers is a signal. If other Chinese models (GLM, Qwen, DeepSeek) follow, the "Chinese vs American models" frontier fades in the cloud. To watch: Microsoft's official policy on Chinese workloads, reaction of U.S. and European regulators, first European enterprise customer running Azure+GLM.
100,000 agents created in 5 weeks by non-developers at the Pentagon: the barrier to building agents drops to zero
The Pentagon built more than 100,000 agents in five weeks through "vibe-coding" (describing what you want in natural language). If this pattern spreads into education, healthcare or local government, the profile "AI user" becomes "agent creator". For Agent Wealthy, this validates the target market: people who never coded can now launch their own automated micro-SaaS.
📊 Trend
July 3, 2026 marks a milestone: the AI agent becomes a layer of national infrastructure. The Pentagon adopts it at scale to support military decision-making, Microsoft commoditizes it via Azure, SpaceX turns it into a $27-billion-a-year financial asset, and security — long the poor cousin — becomes a full-fledged open-source product with Exabeam and Ory. Consequence for anyone building with AI: a serious agent in 2026 can no longer ship without an identity framework, a behaviour-monitoring layer, and an adversarial pre-deployment test. The weekend prototype becomes critical software that demands the same safeguards as a banking system.