Latest briefing
July 16, 2026 · 5 stories (site) · 7 stories (base)
On July 16, 2026, AI agents move to a new phase — the one where they stop being prototypes and become operators of critical processes. Oracle opens a workshop to create them directly inside its ERP, at the very heart of the software that runs large enterprises. Anthropic ships a native browser inside Claude Code, so the coding agent sees, clicks and fixes its own work without human intervention. Lyzr raises $100 million by letting its own agent run the negotiation — large-scale proof that an agent can hold a high-value financial role. On the open-model side, Qwen passes one billion downloads on Hugging Face and overtakes Meta Llama — the centre of gravity of open AI shifts towards China. And Apple sues OpenAI for alleged theft of secrets about the next consumer device, the first large-scale test of trade-secret protection in an era when talent moves freely.
🔥 Top story
01
Oracle opens a workshop to build teams of AI agents directly inside enterprise management software
Behind every invoice, every order, every salary payment sits a large enterprise software — an ERP — that for decades has mostly just recorded transactions. On July 14, Oracle — which equips tens of thousands of large enterprises — opened a new workshop (Oracle AI Agent Studio) that lets anyone create applications made of specialised AI agents directly inside its ERP. Three doors in: dictate what you want in plain English to the computer, snap together pre-built blocks, or write the code yourself in VS Code and let your own coding assistant (Claude Code, OpenAI Codex) handle it. The agents created this way run inside Oracle's engine and automatically inherit its security rules, history and controls — in practice, your ERP learns to do the work itself rather than just note it down. More than 80,000 experts are already trained on this workshop. The stakes for IT departments: moving from a system that takes notes to a system that executes.
02
Claude Code ships a native browser: the agent sees, clicks and fixes its own code by itself
When a developer asks an AI assistant to code a sign-up page, you usually still had to open Chrome yourself to check that the button works, that the form responds, that the colour passes. Anthropic added, on July 10, a browser built directly into its Claude Code tool on Mac and PC. The agent opens the documentation, launches the test server, clicks its buttons, fills in forms, takes screenshots — and loops until it works. In practice, it's like an intern who never sleeps and never asks you anything. Strict guardrails: the agent is not allowed to place an order, create an account, or bypass a CAPTCHA without your explicit permission. First major vendor to integrate this natively into a tool for code-by-agents.
03
A startup raises $100 million by letting its own agent run the negotiation
The deal is one of the most unusual of the year. Lyzr, an American startup that builds AI agents for large enterprises, closed on July 9 a $100 million funding round — about $500 million valuation — with an unexpected operator at the wheel: its own in-house agent, SivaClaw. This agent answered more than 130 investor questions, drafted the pitch documents, and watched the time each candidate spent on each slide to gauge their interest. In total, $400 million of interest received, without any founder getting on a plane to Silicon Valley. Large-scale proof that an agent can run a high-value financial process — and a signal that agents are no longer assistants but full operational colleagues.
04
Qwen passes one billion downloads and becomes the world's most popular open AI family
Alibaba — the Chinese e-commerce giant — just passed one billion cumulative downloads of its Qwen family on Hugging Face, the large library of AI models where developers worldwide come to pick from. Qwen now overtakes Meta Llama, long the leader, and becomes the most downloaded open-source AI model family in the world. Three reasons explain this shift: Qwen is designed from the ground up to work well in languages other than English — French included —, it comes in several sizes adapted to every kind of hardware from a small PC to a big server, and its licence (Apache 2.0) allows any commercial use without cap. According to the Spring 2026 Hugging Face report, Chinese labs now account for about 41% of global downloads — a major shift in the centre of gravity of open AI, leaving the American West Coast for Beijing.
05
Apple sues OpenAI over alleged theft of secrets about its next consumer device
On July 10, Apple filed suit in a California federal court against OpenAI, two of its former employees now executives at OpenAI, and its hardware subsidiary acquired last year. Apple accuses OpenAI of deliberately obtaining secrets about physical-device design — engineering methods, supplier choices, manufacturing techniques — via two transfers: Tang Tan (former Apple VP of Design, now hardware chief at OpenAI) and Chang Liu (former systems engineer, who kept an Apple laptop connected to Apple servers after leaving). More than 400 former Apple employees now work at OpenAI. The stakes go beyond the trial: the consumer device launch expected in 2026, designed by OpenAI with Jony Ive (ex-Apple), could be blocked or slowed by an injunction. First major judicial confrontation between two tech empires on physical AI — and the first large-scale test of trade-secret protection in an era when talent moves freely.
📡 To watch
Mistral teases a 'summer' model expected this week — direct rival to GLM-5.2 and Kimi K2.6
Arthur Mensch (CEO Mistral) confirmed in early July that a new 'summer' model would be available in early access during July, positioned as a direct rival to GLM-5.2, Kimi K2.6 and DeepSeek V4. Apache 2.0 licence expected, in the European open-weight DNA. Watch in the next 7 days: actual release, independent benchmarks, parameter size, early-access conditions.
RASA threatens to push international customers off American clouds — towards Huawei and AliCloud
The Remote Access Security Act (RASA), filed simultaneously in the US House and Senate on July 13, would require cloud operators worldwide to ask each customer to prove they are not a Chinese user before accessing American AI models. Stanford analysis: counter-productive risk of pushing legitimate customers to Huawei Cloud and AliCloud. Vote to follow — indirect impact on agent workloads running on AWS / Azure / GCP from Norway.
The Apple vs OpenAI lawsuit could force OpenAI to reveal who knew what about the Jony Ive device
If the court orders OpenAI to identify precisely which teams saw which Apple trade secrets, the 2026 hardware launch with Jony Ive could be delayed, redesigned, or blocked. Possibility of an injunction blocking possession, use or disclosure of the information. To follow: preliminary hearings, sealing orders, official OpenAI statements on the product schedule.
Microsoft Agent Framework GA 1.0 — first production case studies now in production
Confirmed in the previous watch (15 July): Microsoft Agent Framework went GA 1.0. Monitoring the first published enterprise case studies (banking, healthcare, retail) — a valuable signal on the move from promise to operational reality.
📊 Trend
On July 16, 2026, the AI agent market moves into a new phase — the one where agents stop being prototypes and become operators of critical processes. Oracle opens an agent workshop inside its ERP, at the heart of the software that runs large enterprises. Anthropic ships a native browser in Claude Code: the coding agent sees, clicks and fixes its own work without human intervention. Lyzr raises $100 million by letting its own agent run the negotiation — large-scale proof that an agent can hold a high-value financial role. On open-model side, Qwen passes one billion downloads on Hugging Face and overtakes Meta Llama — the centre of gravity of open AI shifts to China. And Apple sues OpenAI over alleged theft of secrets about the next consumer device, the first large-scale test of trade-secret protection in an era when talent moves freely. For those building with AI, three lessons emerge: (1) agents move out of technical demos and settle into the core of business software — the question is no longer 'can we?' but 'how fast?'; (2) agents become end-to-end operators of financial, logistics and HR processes — not assistants; (3) the choice of underlying model becomes a geopolitical choice — Chinese Qwen or US frontier, with performance / cost / sovereignty trade-offs to make for each use case.