Latest briefing
July 18, 2026 · 5 stories (site) · 6 stories (base)
On July 18, 2026, the AI agent market keeps industrializing and hardening: Thinking Machines Lab releases Inkling, an open model designed from the ground up to drive agents; PrismML compresses Bonsai 27B down to the point where it runs directly on a high-end phone; Tec-Do launches Navos 2.0, a complete chain of marketing agents covering everything from opportunity detection to publication; Alterion positions Draco as the control layer no enterprise deployment can do without; and Bunkerhill Health reaches $55M in cumulative funding for agents already running in American hospitals.
🔥 À la une
01
An open, free model that understands text, images and audio — but only on very large machines
Behind every AI assistant sits a « model » — a huge file of numbers trained to recognize and produce language. Thinking Machines Lab, a young startup founded by former OpenAI staff, has just released Inkling under an open license: one of those models, capable of reading text, images and audio, and designed from the ground up to drive agents that use tools or write code. The commercial bet is straightforward: Inkling isn't necessarily the best on every benchmark, but it is open, modifiable, and shipped with documentation that explains how to adapt it to a specific trade. It's the foundation an enterprise can take over to build its own tailor-made agent. The catch: even the compact version still needs around 600 GB of video memory — the equivalent of half a dozen high-end graphics cards. In practice, this model won't run on a typical SME server; you'll either go through a cloud provider or plan for a lighter version. The lesson is clear: openness is becoming a sales argument, but hosting remains a cost line to compare on a per-task basis.
02
A full AI agent that fits in a high-end phone's pocket
Running a full artificial intelligence directly on a phone, without sending any data to the cloud — that's what Bonsai 27B promises. The American startup PrismML has compressed a 27-billion-parameter model — equivalent, in complexity, to an encyclopedia of several thousand volumes — down to a file of barely 3.9 gigabytes. Small enough to fit in the memory of a recent iPhone. On paper, the model accepts text and images, can reason, call external tools, and respond at about 11 words per second, right on the device. This opens up a new kind of product: a « private agent » that no longer ships your personal data to a remote server, works without an internet connection, and could one day cost less to run than cloud offers. It remains to be checked on real hardware, in real conditions, that these numbers hold up. But the promise changes the game for everything touching confidentiality: an agent that analyzes legal, medical or financial documents without ever leaving your pocket.
03
A chain of agents that runs the entire export marketing process, from idea to publication
Instead of selling « an assistant you can chat with », the Chinese startup Tec-Do unveiled on July 17, at the WAIC trade show in Shanghai, Navos 2.0: a true marketing production line, made up of several specialized agents that hand off work to each other. One scans foreign markets for opportunities. Another writes ad copy. A third identifies the right local influencers. A fourth produces the short videos adapted to each platform. A fifth publishes and adjusts, from your computer, inside the tools you already use (Excel, PowerPoint, Lark). The lesson is direct for anyone who wants to sell agents: the value no longer lies in the conversation with a model, but in a complete, measurable commercial outcome. For a French SME targeting export markets, what you need to sell is « your multilingual campaign ready to publish with its prospects identified » — not « an agent that helps you think ». Tec-Do claims more than 100,000 advertisers served in 2025 and a network of 150 million creators; these numbers come from the company itself and still need validation on concrete cases.
04
A control layer that stops an agent from erasing your data or touching production
Testing an agent before putting it into service is good. Monitoring it while it works is better. That's precisely what Draco offers, launched on July 16 by startup Alterion: a software layer that sits between the agents and the company's systems, watches what they're about to do, and blocks the action if it's risky — deleting data, modifying a production system, an unusual bank transfer. In practice, each agent receives limited permissions, its actions are logged, and a human can validate sensitive decisions with a single click. The product also promises to measure the « token » consumption (the units billed by AI providers) of each agent, so you know who costs what. For IT directors, this is the missing brick: without this kind of safeguard, deploying an agent in a critical system is like hiring someone and handing them the keys without checking what they do. It is also, for offers like Agent Wealthy, a chance to sell an « enterprise-ready agent » product, with security and audit built in from day one.
05
$55 million for agents already working in American hospitals
A platform that lets hospitals create and deploy their own agents — for patient triage, treatment follow-up, administration — has just closed a funding round that brings its cumulative financing to $55 million. Bunkerhill Health, a California startup, convinced prestigious funds (Khosla Ventures, Sequoia, Y Combinator) that AI agents can hold their own in critical medical processes. At the University of Texas Medical Branch, more than twenty agents are already running in parallel on the Carebricks platform. According to the company, a nephrology triage agent has cut the time to see a specialist in half; a lung nodule monitoring agent handled urgent cases 80% faster and doubled the number of patients followed up according to recommendations. These results are reported by the company and remain to be confirmed on other sites. But the underlying trend is clear: investors are betting on vertical agents — specialized in a specific trade, with a measurable and reproducible result — rather than on general-purpose versatile agents.
📡 À surveiller
$13M for a startup that cleans data before handing it to cybersecurity agents
Beacon Security closed a $13 million seed round on July 16. Its specialty: taking information from dozens of different security tools (firewalls, antivirus, intrusion detectors), cleaning it, cross-referencing it, then handing it to agents tasked with detecting attacks or investigating. The message applies to any agent: if it receives incomplete or contradictory data, it makes bad decisions. On a site like The Agent Watch, a useful answer must be traceable to its source and separate verified data from a simple claim.
Claude Code turns a conversation into a background task that keeps working without you
Anthropic added the « /fork » command to Claude Code: it copies the conversation into a separate task that keeps working while the user moves on. Pages produced by Claude can also query the reader's own tools (for example their GitHub list), with their consent. The published deliverable thus gets closer to a living mini-product rather than a frozen snapshot.
NVIDIA proposes to measure agents in « intelligence per dollar » rather than in tokens
NVIDIA considers that agents must be continuously improved from their failures in production, and recommends measuring what this loop actually costs. The notion of « intelligence per dollar » invites tracking, for each agent, the full cost per accepted deliverable and its success rate, rather than just the price per million tokens.
Verifiable data and provenance: a standard to enforce from day one
Beacon Security shows it: an agent fed on unverified data makes bad decisions. For any agent offering, every response must indicate where the information comes from, separate verified data from a vendor's claim, and keep a trace of the source. It's also what a briefing like The Agent Watch strives to do: cite the primary source for each story.
📊 Tendance
On July 18, 2026, the AI agent market keeps industrializing and hardening. Two open models stand out: Thinking Machines Lab's Inkling, designed from the ground up to drive custom agents, and PrismML's Bonsai 27B, compact enough to run a multimodal agent directly on a phone. On the product side, Tec-Do launches Navos 2.0, a complete chain of marketing agents covering everything from opportunity detection to publication, while Alterion positions Draco as the security layer no enterprise deployment can do without. Money follows the same logic: Bunkerhill Health ($55M cumulative) proves investors now fund vertical agents that deliver reproducible results in hospitals, while Beacon Security ($13M seed) bets on the quality of data fed to cybersecurity agents. Three lessons emerge for those who build with AI: (1) a sellable agent is now an agent that combines a model fitted to the use case, a complete journey from need to outcome, and a security and audit layer; (2) open models become a commercial edge — modifying the engine is more profitable than reconfiguring everything through APIs; (3) the right metric is not the price of the token but the full cost per accepted deliverable and its success rate, as NVIDIA proposes.