The Agent Watch
Briefing · Articles · Tools · About EN FR DE ES 中文 IT PT SV FI DA

Daily Briefing

16 November 2026 · 5 stories

🔥 Top story

1

OpenBot 1.0: the agent works on your machine, but it has to ask permission

CopilotKit releases OpenBot 1.0, a workshop where you install agents that do the work for you. Each agent gets its own computer: a browser you can watch working, files, and a command line. You give it three kinds of rights: it may act on its own, it must ask your agreement, or it is not allowed to touch it at all. Every attempt is written in a log. The example given by the publisher is an agent that checks expense claims. It looks up the rule in force, shows the receipt line that raises a question, and flags what needs human judgment. It is not allowed to trigger the refund. The code is free, under an MIT license, and the control room that provides memory and governance is sold separately. Source : https://www.copilotkit.ai/blog/introducing-openbot-1-0

2

Stanford teaches a model to keep just 16 memories of a whole library

Researchers at Stanford open the code of a memory module for frozen models, under an Apache 2.0 license. The principle: the model no longer reads every document for every question. It keeps only 16 tokens, chosen according to the question asked, as if you kept 16 note cards in a card box instead of dragging a whole library around. On a factual question benchmark, it reaches 84.33 % with those 16 tokens, against 80.43 % when it is given everything. The number of document tokens is 108 times lower. On 512 documents, the measured latency drops from 19.9 seconds to 0.74 second, 27 times less. The authors also state the limits: the module loses on three benchmarks, when the documents are too short for compression to gain anything. Source : https://memory.stanford.edu/

3

Macs rented by the day for agents that need a real computer to work

Mount Thor raises 45 million dollars to sell Macs on rental, the way servers are rented. The idea comes from a simple observation: an agent that drives a computer needs the same environment as a human, with its applications, its files and its screen. Many tasks exist only on Apple's operating system. The company therefore sells complete machines and isolated virtual machines, billed by the day. It has already opened its first data center in less than six months. Among the investors are well-known names from the sector: Meta, OpenAI and Vercel, alongside CRV and Conviction who lead the Series A. The team previously built infrastructure at Cerebras, CoreWeave, OpenAI and xAI. Source : https://blog.mountthor.com/a

4

Kindo raises $30 million to track what every agent costs

Kindo raises 30 million dollars, co-led by Crosspoint Capital and Forgepoint Capital, for a total of 57.6 million since it was founded. The company sells a control room for company agents: knowing what they do, what they touch and what they cost. The most concrete part is the bill. The console records every token consumed and the result of every run, including the failures, then rolls the costs up by person, by agent, by service and by project. Finance teams can set hard ceilings and trigger alerts. Kindo says it grew revenue by about 800 % in a year, and cites a customer who says it saves more than 2 million dollars a year. These are publisher figures, not audited. The product works with any agent, even one built elsewhere, and can run on the customer's own site, with no connection. Source : https://www.kindo.ai/resources/press-kindo-raises-57-million-ai-agent-management

5

Blue J raises $176 million for agents that do tax work, not just answers

Blue J raises 176 million dollars in a Series E round led by JMI Equity, and is valued at more than 700 million dollars. The company sells an agent that does the tax professional's work, not only the one that answers questions. You hand it a whole task: reviewing a return before filing, preparing a reply to a tax notice, analyzing a partnership agreement, drafting a letter to the client. It produces a document ready to be reviewed, with its sources and its steps available for inspection, and the human keeps the judgment. The platform has already answered more than 14 million tax questions. Annual recurring revenue grew 165 % in fifteen months. More than 6 500 organizations and 45 000 professionals use it. The founder sums it up: an answer is almost never the end of the client's work. Source : https://www.bluej.com/blog/blue-j-announces-series-e-funding

📡 To watch

Solo.io wants an agent session to pause instead of keeping a server busy

Solo.io announces the enterprise versions of two pieces for running agents in production. The idea: an agent session pauses when it has nothing to do and resumes on demand, instead of leaving an empty server running all the time. Each session is isolated, and a gateway checks who calls it and what it can reach. OutSystems and Swissquote already use it. Source : https://www.solo.io/press-releases/announcing-kagent-1-0-ga

Mirantis adds an emergency stop and spending caps for agents in production

Mirantis puts on sale a platform for running agents on your own servers. It lets you define what each agent can reach, set spending ceilings at three levels, record everything, and cut the network of the whole fleet in one move during an incident. An agent may propose a repair on a production machine, but a human must approve it before it is applied. The publisher relies on two studies: 49.8 % of the organizations surveyed had an agent take an unwanted action or suspect one, and only 17.5 % of them can stop it automatically. Source : https://www.mirantis.com/company/press-center/company-news/mirantis-announces-agent-platform-that-strengthens-human-oversight-of-production-ai-agents

AT&T runs nearly 500 processes with agents, without dropping the human too early

The telecom operator AT&T has put nearly 500 processes managed by agents into production, and plans to triple that number in 2027. Its platform handles more than one million transactions a month. The interesting point is not the figure but the rule: every process starts with a human who validates, and full autonomy only arrives once accuracy reaches the mid-90s and the risk is judged acceptable. An internal group, Business Agent Foundry, maps the processes and sends the most sensitive ones through security and compliance reviews. Source : https://siliconangle.com/2026/11/14/att-uses-agentic-ai-to-reinvent-business-processes-from-the-ground-up

A voice model that knows who was talking, even when the television answers

Bland presents Tofu-1, a model that listens and speaks at the same time, without going through transcription. The publisher targets hard calls: from a moving car, in a hospital ward, with several people talking. Its favorite test is the television case. The agent asks whether the caller wants to cancel, the caller does not answer because they are looking for their contract number, and the television says yes. A model that reads only text would take that yes for a confirmation. Tofu-1 is designed to attribute the sentence to the television and ask again. Access is a research preview, for a few companies. Source : https://www.bland.ai/speech-to-speech

📊 Trend

Today, agents are installed and controlled as much as they answer. OpenBot 1.0 gives each agent its own computer, its rights and its log, and Stanford shows that a frozen model can hold a whole collection of documents with 16 memories, which makes the cost predictable whatever the volume. Meanwhile the invoice becomes a management topic: Kindo rolls up what each agent costs, Mirantis lets you cut the network of a whole fleet in one click, and Solo.io wants an agent session to pause instead of keeping a server busy forever. AT&T applies the same rule at scale: a human validates at the start, and autonomy only comes after a measured accuracy. Finally, the money moves toward the business application: Blue J sells an agent that does the tax work, not one more answer.