Daily Briefing
27 July 2026 · 5 stories
🔥 Top story
01Visa and Lianlian: an AI agent buys and pays on its own in B2B
For the first time in mainland China, an AI agent actually found a supplier, compared prices, placed an order and settled the invoice, all in one go. The operation, run by Lianlian's LoopXPay, was framed by pre-set spending limits and authorisations. The agent was also registered in Visa's official directory, which acts as a verifiable ID. Think of it as a meticulous assistant handed your corporate card, with a strict rulebook: it can commit to the purchase, but never without a guardrail. For businesses, this opens the door to supplier sourcing and cross-border payments fully run by software.
02Siemens and NVIDIA: an agent that double-checks its own work before delivering
Siemens just made its electronic design agent far more reliable: instead of proposing a solution and waiting for a human to approve it, it continuously verifies its decisions inside physics-based simulation software. The NVIDIA building blocks include reasoning models, a secure execution space and high-performance compute. The reported outcome: ten times faster on some chip design steps, five to ten times less digital fuel burned. Imagine an engineer who, after every decision, automatically re-runs the simulation on a dedicated machine. For industry, this is the shift from an 'assistant that suggests' to an agent whose every step is verifiable.
03Gemini Intelligence: an agent that acts across 40 apps on your phone
On the new Samsung foldables, Google's agent can chain actions across more than 40 apps: book a restaurant, order a ride, buy a ticket, track an order. It works in the background but keeps you in the loop before any sensitive action. You can show it a menu photo and ask it to book a table, then follow each step as it runs. For now, the service is limited to adult users in some countries, languages and markets. This is the first truly OS-level agent deployment for the general public, like a digital colleague who already knows your apps.
04Natural raises $30 million: bankers for AI agents
Natural, a US startup less than a year old, has just raised $30 million in Series A to give AI agents the financial tools they lack: wallets, accounts, fraud detection, identity verification. Six products are already live, seven more are on the roadmap. In practice, an agent will soon be able to hold money, pay a bill or be turned into a merchant, without a human holding the card. Think of it as opening a bank account for a digital employee, with restrictions, audit trails and proper controls. For investors, it's the signal that the financial layer of agents is now a real business.
05Kimi K3: the long-awaited Chinese model still has to show up
Moonshot has been promising for weeks to release the open weights of its Kimi K3 model today, 27 July. At 06:00 Paris time, the official page still showed no file and no final licence. The model is billed as one of the largest publicly open models ever: close to 3 trillion parameters, text and image, a one-million-token window. But an independent test flagged hallucinations in 51% of cases. Everything is ready for the demo — the question is verification: the promise will be delivered today, or make headlines again tomorrow.
📡 To watch
Sunrate and Mastercard publish an industry framework for agent payments
The white paper lists 16 blockers and 13 concrete use cases, from supplier to compliance.
Aiven buys Flow AI: the infrastructure layer for agents consolidates
The deal brings analytics agents closer to real enterprise data. Amount undisclosed.
Shoplazza's Athena: an agent to run an entire online store
The public launch widens visibility of a product announced in May. Catalogue, orders, logistics and ads are under agent control.
Kimi K3: actual licence and real-world conditions
Beyond release, terms of use and required hardware will decide who can really run it.
📊 Trend
AI agents are moving from prototype to production: they buy, design circuits, book trips or run a store. Every new announcement highlights the same shift: for an agent to be useful, it needs verifiable rules, a recognised identity and a dedicated financial layer. 2026 will be the year the rails of agentic commerce become a real business.