Daily Briefing
14 November 2026 · 5 stories
🔥 Top story
1ChatGPT now has a wallet: the agent can pay, but never without you
OpenAI adds a wallet to ChatGPT, so your assistant can now settle an order on your behalf. The principle fits in one sentence: the agent pays, but it never sees your card number. You register your card once, then approve every purchase with a fingerprint or a face scan, the way you unlock your phone. The shop only receives a one-time token with no link to a real bank account. You keep control of prices, returns and disputes. The service is limited to ChatGPT Pro subscribers, in the United States only. The next announced step goes further: permanent delegation with a spending cap, allowed categories and time windows. Source : https://openai.com/index/introducing-chatgpt-wallet
2Sierra freezes the way your agents introduce themselves to shops
Sierra releases version 1.0 of Poppy, an open standard that explains to a personal assistant how to introduce itself to a business. In practice, it is a digital ID card for agents: it says who the agent is, what it is allowed to do, and the customer can withdraw those permissions at any time. Shops in more than 175 countries that use the Shopify platform switch it on automatically, with nothing to install, and Meta's Muse assistant uses it everywhere. More than one hundred companies are already testing the standard. The shop keeps control of prices, customers and returns: the agent walks through the door, it does not take the keys. Source : https://sierra.ai/blog/releasing-poppy
3Google measures its guardrail: 87% of harmful instructions executed are flagged
Google publishes the results of its anomaly detector, a guardrail that reviews afterwards what agents did. On 1,774 test sessions, drawn from five test suites written by others, the system flags 87% of the harmful instructions agents actually carried out. It is formidable against goal hijacking and wrong tool use. On the other hand it is almost blind when the instruction hides in the agent's long-term memory: only 4 cases out of 18. In other words, the alarm works when the request is visible, and stays quiet when it is slipped into the data. Source : https://developers.googleblog.com/how-agent-anomaly-detection-scores-on-benchmarks-we-didnt-write/
4LangChain raises $300 million to keep agents inside companies
LangChain announces a $300 million round that lifts its valuation to $3 billion. The round is led by Google's growth fund, with Singapore's sovereign fund and DFJ Growth as new investors. The company says more than 7,000 organizations use its paid platform, including Nvidia, Toyota and Salesforce. Its bet fits in one sentence: a company that controls its own agents also keeps what they learn, and that know-how becomes hard to copy. It is the direct counterweight to the offers of the big model and cloud providers. Source : https://www.langchain.com/blog/series-c
5A 145 billion parameter model fits on a single graphics card
Edge0, an American lab, releases the weights of its Edge0-145B model: 145 billion parameters in total, of which 10 billion work on every word. It fits on a single 24 gigabyte graphics card, in an ordinary desktop computer. The publisher claims 121 words per second on agent work and 86.5 points on a test measuring an agent's work on a terminal. The files download without an account, in two formats. The licence builds on Qwen's and requires showing the model name beyond 100 million monthly users or $20 million in monthly revenue. These figures come from the publisher, with no independent evaluation known. Source : https://huggingface.co/Edge0/Edge0-145B-q3k
📡 To watch
NiCE ships a new generation of its agent platform, with a log of every decision
Customer relationship software maker NiCE releases the new generation of its Cognigy platform. Three things stand out: a companion that guides teams through building and testing agents, skills that are reused from one agent to the next, and a log that retraces every decision, every tool called and every token spent. The platform also talks to customers' own agents through open standards, and handles speech end to end. Existing customers keep what is already running. Source : https://www.nice.com/press-releases/nice-launches-next-generation-nice-cognigy-to-accelerate-enterprise-ai-agent-deployment-and-orchestration
Google Cloud puts agent monitoring into company dashboards
Google Cloud adds agent monitoring and control to its data tools. Every agent can send its usage figures, response times and tokens consumed into tables the company owns, then analyse them in the tools it already uses. Full conversations are optional: by default you see the activity without seeing the content of exchanges, and the detailed log is switched on agent by agent, with its guardrails. Prebuilt dashboards track active sessions and the most consulted sources. Source : https://cloud.google.com/blog/products/data-analytics/enterprise-ai-agent-observability-with-bigquery-and-looker
Amazon shows how to pay an agent by the result, not by the word
Amazon publishes a guide to send each request from a coding agent to the most suitable model, based on the failures recorded in the history. The guide also proposes the right unit of measure: cost per solved task, rather than price per word. A more expensive model that gets there can cost less overall. Amazon notes that the example is demonstration code, with no production guarantee. Source : https://aws.amazon.com/blogs/machine-learning/route-coding-agent-requests-to-the-right-model-on-amazon-bedrock-with-nvidia-nemo-switchyard/
llama.cpp reaches version 1.0 and promises to stop breaking your agents
The free software that runs models on an ordinary computer reaches version 1.0, after three and a half years. The team commits to no longer changing the functions used by programs without changing the major version number. More than 150 model architectures are recognised, and the files exist for Linux, macOS, Windows, Android and iOS. For companies, that stability promise is the decisive argument: an agent running locally no longer breaks at every engine update. Source : https://huggingface.co/blog/ggml-org/llama-cpp-v1
📊 Trend
A day when the agent moves to the money and control side: OpenAI opens a wallet for it, Sierra gives it an ID card to enter shops, Google measures what its guardrail is worth, and LangChain raises $300 million on the idea that companies will keep their agents in house. On the other side, the tools settle down: a giant model fits on a desktop graphics card, the free software llama.cpp promises to break nothing any more, and Amazon learns to pay an agent by the result. The agent becomes a colleague you equip, control and invoice.