Where do AI agents actually go to live? Today gave two answers. Meta opened its Business Agent Platform worldwide, dropping agents straight into the WhatsApp, Messenger and Instagram threads people already use, while NVIDIA and ServiceNow built one that sits on a worker's desktop and remembers them from one day to the next. Around those launches, OpenAI and Google made their existing tools a little roomier, Unitree turned heads with a 650,000 dollar mecha robot, and Anthropic quietly strengthened its bench. No new frontier models shipped, but the direction of travel is clear: the contest has moved from building bigger models to getting agents in front of the people who will actually use them.
Meta opens its Business Agent Platform worldwide and starts renting out its own compute
If you build customer-facing tools, this is the release to watch this week. Meta has switched on a global Business Agent Platform that lets companies build, customise and deploy AI agents straight into the billion-plus conversations already flowing through WhatsApp, Messenger and Instagram, a distribution route none of its rivals can match. Alongside it, the new Meta Compute business rents out the company's spare data-centre capacity, putting Meta into direct competition with AWS, Google Cloud and Azure.
The practical takeaway is that the hardest part of enterprise AI, getting an agent in front of customers who already trust the surface, is exactly what Meta is now selling as a service. For anyone weighing agent platforms, it is worth comparing this against Google's Gemini Enterprise and Microsoft's Frontier Company before committing.
NVIDIA and ServiceNow launch Project Arc, a self-evolving desktop agent
This matters if you have been frustrated by assistants that forget everything the moment you close the tab. Project Arc is a long-running desktop agent that persists on a knowledge worker's machine, learns their workflows over days and weeks, and keeps improving, rather than answering one prompt and resetting. It runs on NVIDIA's OpenShell secure runtime with open Nemotron models, which is the pitch that you can have a persistent agent without shipping your data off to a frontier lab.
The ServiceNow pairing puts it right inside the IT and operations backbone where many large companies already work, so it has a natural home. The real test is reliability over weeks rather than a polished demo, but if it holds up, the always-on agent starts to look like the default enterprise shape and the stateless chatbot starts to look dated.
ChatGPT custom instructions now stretch to 5,000 characters
If you have ever hit the wall trying to tell ChatGPT how you want it to behave, this is a quietly useful change. OpenAI has raised the custom instructions limit from 1,500 to 5,000 characters, giving far more room to spell out tone, format, context and standing preferences that persist across chats. The catch is that the larger allowance is reserved for paid tiers, so Pro, Enterprise, Business and Education users get it while free accounts stay on the old limit.
For heavy users running repeatable workflows, a longer instruction block means less repeated prompting and more consistent output, which adds up over a working week. OpenAI also published updated prompting guidance to help people use the extra space well rather than padding it out. Treat it as a reason to revisit and tighten your saved instructions rather than simply making them longer.
Higher Gemini usage limits land in Google Sheets
For anyone who leans on Gemini inside a spreadsheet, the ceiling just went up. From today, accounts with AI Expanded Access licences get higher limits on Fill with Gemini and the AI function in Sheets, which means fewer interruptions when you are running Gemini across large ranges of data. The change is incremental rather than headline-grabbing, but it removes a real friction point for people using AI to clean, classify or summarise rows at scale. If your team has been rationing these calls to stay under the old caps, it is worth checking whether your licence now clears the higher tier.
ATL Saathi, a Gemini-powered web app, is now piloting in 100 Indian schools, giving Tinkering Lab educators a round-the-clock planning assistant with micro-learning, project ideas, multilingual support and curriculum-aligned guidance. Separately, DeepMind, Faculty and the UK government are trialling a Gemini-powered planning assistant to help councils clear administrative backlogs, aiming to halve the manual work behind household planning applications. Google DeepMind, 15 Jul →
Unitree unveils the GD01, a 650,000 dollar transforming mecha robot
This one is more spectacle than shopping list, but it signals where embodied AI is heading. Unitree, known for affordable humanoids and robot dogs, has revealed the GD01, a giant transforming robot priced at 650,000 dollars and timed to its roughly 619 million dollar Shanghai IPO. It is a halo product rather than a practical tool, designed to show the company can build at the high end of scale and spectacle as well as the low end of cost.
For anyone tracking robotics rather than pure software, the interesting part is the strategy: using a viral flagship to punctuate an IPO week and broaden perception of what the company can do. It tells us little about whether humanoid robots make economic sense at scale, which remains the open question over the whole sector, but it keeps Unitree firmly in the conversation.
Anthropic has reportedly added Andrej Karpathy, the former Tesla AI director and OpenAI founding member, and Monzo co-founder Tom Blomfield, who joins the AI compute team. It is a talent story rather than a product one, but it is a signal worth filing, since where respected practitioners land tends to shape model quality and shipping pace a year out. Build Fast with AI, 15 Jul →
Today's themes
The clear pattern of the day is that the enterprise agent, not the chatbot, is where the serious effort now sits. Meta's Business Agent Platform and the NVIDIA and ServiceNow Project Arc both landed within hours of each other, and both surfaced through the companies' own channels before mainstream coverage caught up, which is exactly why the social scan runs first. The second theme is that the real movement this cycle is in access, limits and distribution rather than raw model releases, with a live tracker showing zero new frontier models in the past 24 hours while OpenAI and Google both shipped quiet but practical capacity and instruction upgrades. The third, quieter theme is anticipation: with Gemini 3.5 Pro widely expected on 17 July, much of the field looks like it is holding position ahead of the next big model moment rather than firing early.