Daily briefing
July 23, 2026 · 5 stories (site) · 5 stories (base)
On July 23, 2026, AI agents are moving from answering to acting: Ushur wants to finish customer processes end to end, NVIDIA brings them as close as possible to robots and creative tools, Fay targets non-technical teams, Washington is looking at Chinese engines, and PsiBot is funding the brains of robots.
🔥 Top stories
01
Ushur launches agents that don't just reply: they complete the request
A customer reporting a claim or updating their coverage should no longer be bounced from one service to another. Ushur is launching a platform where the agent understands the request, gathers the documents, acts inside the company's software, and guides the person all the way to resolution. It's the difference between a switchboard operator who hands out a phone number and an assistant who actually takes charge of the file. The tool targets insurance, healthcare and banking, where every action must be controlled and auditable. It works over text, email, web or voice, without forcing customers to repeat themselves when they switch channels. Ushur announces more than 200 connections to business software and offers a no-commitment trial. The promise is clear; its effectiveness will now have to be confirmed on real-world journeys.
02
NVIDIA takes agents out of the cloud to run robots and creative tools on site
A robot, a car or a creative studio can't always wait for a reply from a remote computer. At SIGGRAPH, NVIDIA is unveiling Cosmos 3 Edge, a compact model able to interpret text, image, video, sound and action directly on a local machine. It can therefore help a robot make sense of its surroundings even when the network is slow or missing. NVIDIA is adding NeMoClaw, a kit to run an agent on a desktop workstation, and is opening a common language between agents and creative software. Adobe, Blender, Houdini and Unreal are among the tools involved. In practice, an agent will be able to spot a missing texture, prepare several exports or oversee a scene, like a studio assistant who knows every workstation.
03
Fay aims to let business teams build a digital colleague in five days
A support, finance or HR team could describe its work in plain words and get an agent that carries it out inside their usual tools. That's the promise of Fay, launched by PropulsionAI. The agent can hook into Slack, Gmail, Jira or PagerDuty, then run a full procedure instead of merely answering a question. The company offers a one-week trial on the customer's real data: the equivalent of onboarding a new colleague on a specific assignment before deciding to hire them. Fay accepts several AI engines, which limits dependency on any single provider. A word of caution, though: the available information comes from a company press release, and no independent assessment of its reliability on complex tasks has been found.
04
Washington weighs closing the door on Chinese models used by US-based agents
US companies running their agents on Chinese models could one day have to switch engines. According to Axios, the Trump administration is discussing restrictions, while these models are said to account for 46.4% of the volume processed by OpenRouter, compared with 35.7% for US models. It's as if a company learned that the engine powering its entire fleet might soon be banned: replacement is possible, but it costs time and money. The success of Kimi K3 and DeepSeek is feeding the debate. No decision has been signed, however: there is still no presidential order or official rule. This information describes internal deliberations reported by Axios, not a ban in force.
05
PsiBot raises close to $100 million to give robots common sense about the physical world
For a robot to fold laundry or pick up a fragile object, it has to understand the consequences of its movements, not just recognise an image. The young Chinese company PsiBot is building this kind of digital brain and has just raised close to $100 million, at a reported valuation of $1.48 billion. Its models learn about the real world using sensor-equipped gloves and humanoid robots, then are tested in logistics and industry. It's comparable to letting an apprentice rack up thousands of hours in the workshop before letting them work alone. The round, led notably by Chery Automobile and Lens Technology, shows that the agent race is no longer confined to screens: it is moving toward machines capable of acting in our environment.
📡 To watch
Kimi K3 promises to open its weights on July 27
Moonshot has announced that it will soon open its flagship model. It will be a real-world test: developers will be able to check whether its performance justifies the political concerns it is already raising.
The common language between agents and tools is due for a milestone on July 28
The MCP standard is set to finalise the handling of long-running tasks and the display of interfaces within tools. For the public, this means agents able to follow a mission for longer and to show clearly what they're doing.
A Chinese model becomes a security-investigation tool in the United States
GLM-5.2 was used to examine an attack against Hugging Face when other models refused to analyse the malicious commands. It is a reminder of the value of having an engine you can control yourself.
PsiBot targets one million hours of physical experience in 2026
The robot battle is also being fought on the amount of experience accumulated. This data is the equivalent of an apprentice's driving hours, but used to teach machines how to handle the real world.
📊 Trend
Agents are steadily moving beyond simple conversation to take charge of complete processes and act in the physical world. Ushur and Fay are placing them inside customer service and internal functions; NVIDIA is giving them access to creative software and local machines; PsiBot is training them to understand gestures and objects. In parallel, the US debate over Chinese models is a reminder that an agent always depends on its engine: a political decision can disrupt a service as surely as an outage. The priority is now concrete: verify the work done, keep a human in the loop for sensitive decisions, and plan for a backup engine.