Daily Briefing
June 26, 2026 · 5 items (site) · 7 items (base)
🔥 Headlines
01
Baseten raises $1.5 billion to become the plumbing that runs millions of AI agents
Picture the plumbing of an office building, but for digital brains. Baseten is one of the few companies that runs the small, specialised AI models that enterprises train for their own agents, in production. On June 22, the company closed a $1.5 billion funding round in two tranches, at $11B then $13B valuation. Concretely, its platform now processes more than one billion calls per day (agents asking the model for a reply), spread across 87 clusters in 18 different cloud providers. To put it in scale: when you send a message to an AI agent, there's a good chance the reply travels through Baseten. The bet is simple: in the age of agents, 30 to 50% of a company's « model » budget will go to custom models — not GPT or Claude — and someone needs to run them at scale. That's the seat Baseten wants to keep.
02
EDB releases a self-healing database designed to serve AI agents
When you ask a conversational agent a question, it goes looking in a database. The problem: agents work non-stop, at full speed, and traditional databases can't keep up. EnterpriseDB (EDB) announced on June 23 a new « agentic » version of its Postgres database that monitors and heals itself, without human intervention. The promise: tuning up to 10 times faster and analytics costs cut by 58%. The guiding idea is almost philosophical: in the age of agents, intelligence must come to the data, not the other way around — done with the days of copying data into a separate warehouse (« data lake ») to analyse it elsewhere. For a company, this is the promise of agents that answer in one second on live data, instead of waiting for a human to copy everything into another system every night.
03
Alibaba open-sources a « reality simulator » to train AI agents
Before letting an agent touch a real computer, you'd want it to train on a copy. That's exactly what Qwen-AgentWorld offers — the open-source model released on June 24 by Alibaba's Qwen team. The system learns to predict what a terminal, a web browser, an Android phone, or a code editor will respond — the inverse of a classic agent that learns to choose the next action. Score on the AgentWorldBench benchmark: 58.71, ahead of GPT-5.4. The model, Apache 2.0 licensed, is downloadable on Hugging Face and GitHub, and runs with 35 billion parameters of which only 3 are active at a time — much lighter than a large general model. Think of a pilot training on a flight simulator before touching the real cockpit: Qwen-AgentWorld is the simulator for agents that type commands on your machine.
04
F5 launches a « guard » to watch what AI agents do inside your company
The application-security giant F5 (NASDAQ-listed) announced on June 22 a complete platform dedicated to AI-agent security, paired with the acquisition of SurePath AI, a specialist in detecting « shadow AI » — those AI tools employees use without telling the IT department. The finding is alarming: 88% of organisations report having an AI-related security issue, and 98% are preparing for the arrival of agents. The platform covers five pillars: discovery of hidden AIs, testing against 140,000 attack scenarios, governance, real-time protection, and observability. It's an antivirus, but designed for software that makes decisions on your behalf — exactly the kind of safeguard IT leaders have been asking for before letting agents touch customer data.
05
SpyCloud ships an agent that runs the cyber investigation instead of the analyst
When a cyber analyst tracks a criminal, they spend hours cross-referencing databases of stolen credentials, passwords, and email addresses. SpyCloud (Austin, Texas) announced on June 24 a conversational agent that does this work in minutes. The agent taps into more than one billion recaptured credentials (infostealer logs, phishing kits, combolists) and automatically reconstructs a suspect's full digital identity profile. Announced results: 8 times more identity records found, 14 times more plaintext passwords, and 5 times more linked emails than a traditional search. Every result cites the records that back it up — no invented answers. The idea: replicate the tradecraft of the best analysts (some former federal intelligence officers) and put it within reach of a junior joining the team.
📡 To Watch
Assort Health becomes a unicorn at $1.2B with its voice agents managing patients
Assort Health closed a $120M Series C on June 24 at a $1.2B valuation. Its platform puts conversational agents on the phone in medical practices: appointment booking, prescription refills, insurance verification, payment collection. 190 million patient calls already handled, revenue multiplied by 20 in 15 months. Healthcare is becoming one of the first sectors where « regular people » talk to an AI agent without realising it.
Labviva ships a crew of mini agents led by a chief for pharmaceutical procurement
In heavily regulated industries like pharma, general-purpose models (GPT, Claude) are expensive and stumble on specialised scientific references. Labviva unveiled on June 24 « The Agentic Crew »: several small models trained on each customer's specific context, supervised by a large model that orchestrates the whole. The result: automating the full « sourcing → buying → payment » cycle with the precision and cost profile the sector demands. It's the archetype of the vertical agent: less brilliant than a generalist, unbeatable on its own turf.
« Your data, your rules »: European sovereignty pushes agents to stay on-premises
As agents touch health, banking or defence data, several players (EDB on the US side, TensorX on the European side last week) lean into the « your AI, your data, your rules » promise. Concretely: agents deployable on the customer's own servers, without depending on an American cloud, and auditable by the regulator. For European customers, this is the precondition for adopting agents in production without falling foul of the law.
Vertical agents by trade are replacing the generalist that knows everything
Labviva (pharma), Assort Health (healthcare), SpyCloud (cyber) and other startups of the moment trace the same trend: the trade-specialised agent takes over from the general model. More precise, cheaper, easier for a CIO or regulator to sign off. The all-powerful « assistant that knows everything » hasn't disappeared, but it becomes the front door; the real workhorses inside a company are agents trained on one specific domain. Like a general practitioner referring you to a specialist — in software form.
📊 Trend
On June 26, 2026, the word « agent » tipped from prototype to infrastructure. Four things converged. (1) The plumbing gets valued: Baseten at $13B shows that running models in production is now worth as much as the models themselves. (2) Data comes to the agent: EDB's « agentic Postgres » embodies the end of the data-lake dogma — AI acts where data lives. (3) Agents train before touching the real world: Alibaba's Qwen-AgentWorld gives them a simulator, like a fake cockpit for a pilot. (4) Agents get watched: F5 and SpyCloud tackle « who guards the agents' room » from two angles — network security on one side, cyber investigation on the other. The story of the day isn't a single product; it's the entire stack being built at the same time: the training ground (Qwen), the data foundation (EDB), the runtime plumbing (Baseten), the guardians (F5), and the agents themselves (SpyCloud, Assort, Labviva). When the whole chain appears at once, the agent economy becomes a real sector — no longer a laboratory experiment.