An AI agent is an AI tool that does a job step by step, instead of only answering one question. I opened every source below myself and read it before writing. When a number comes from the company selling the product, I say so.
Sierra and Meta draft a standard for personal shopping agents
On 6 October, The Next Web reported that Sierra announced the Personal Agent Protocol, an open standard for how personal AI agents sign in to businesses and what they can do there. Sierra and Meta are developing it. Named partners are Genesys, Instinct, Rocket, Shopify, Stripe and Walmart.
Sessions are built on OAuth, a common way to log in. An agent can start as a guest, then get read-only or write access once the customer signs in. A business chooses whether agents deal with its website, its APIs, or its own agent. The first version, v0.1, is due later in October.
Payments, push notifications and finer permissions are left for later. The article points out that European rules on payment approval were written for a person approving a payment, with no exemption for software approving on someone's behalf. It also says no European retailer, bank or payment company has been named. Its line: "The agents are arriving faster than the arrangements they will work under."
My take: Sierra and Meta are writing this standard, so they gain if it wins. I read one source, The Next Web, and the spec itself is not out yet. It is a plan for how agents could shop for you, not something you can use today.
Nous Research raises $90M to take its Hermes agent to businesses
Nous Research raised $90 million to build a business version of Hermes, its open-source AI agent. The Next Web and Dealroom both reported it on 7 October, and both lean on The Wall Street Journal, which first reported the news.
The $1.5 billion valuation comes from the WSJ. Nous did not give a valuation itself. The CEO, Dillon Rolnick, named investors including Nvidia, Microsoft's M12, Samsung, Robot Ventures, Union Square Ventures, Y Combinator and Menlo Ventures.
Hermes has been open source under the MIT licence since February. Nous says it has been cloned more than 24 million times and estimates it drives about 2.5% of global token usage. Those are its own figures. The WSJ reports more than 22 million downloads. According to Nous, the business version will let companies pick models by price and goal, keep AI use private and own their company knowledge. Hermes works through Telegram, Slack and WhatsApp.
Dealroom also repeats a WSJ revenue claim: about $36 million a year by mid-September, with a hope of passing $100 million by the end of 2026. I only saw that in one place, so treat it with care.
My take: Most of the usage numbers here are Nous's own estimates. Some of its backers, such as Microsoft's venture arm and Nvidia, also sell AI products, so they gain if agents take off. Money raised shows interest, not that the business version works well.
Epoch tests whether AI agents can do its own job
On 8 October, Epoch AI introduced its Automation Reports, built from real tasks in its own research work. Crypto Briefing and AlphaSignal both covered it. Six models tried 11 assignments, and Epoch staff graded the results against the standards they use for their own work.
Claude Fable 5.1 and GPT-6 Astra scored highest, at about 65% each according to Crypto Briefing. Grok 4.6 got 59%, Qwen 3.8 Max 53%, Kimi K3 52% and Gemini 3.8 Flash 42%. No model reached Epoch's bar for replacing an employee. Crypto Briefing's wording: "Frontier models get close on clearly defined tasks, but they still fall short of fully automating the work."
The models did better on well-defined work and worse on open-ended, judgment-heavy work. AlphaSignal adds one setup mistake: a 4,096-token output limit cut off 61 of 280 answers in one GPT-6 Astra experiment, yet those answers still counted as attempts. Both sources warn it is only 11 tasks from one organisation, and human grading is subjective.
Disclosure: Anthropic makes Claude, and I use Claude to help run this blog. Read this section with that in mind.
My take: OpenAI, Anthropic and the other model makers did not write this report, Epoch did. The pages I read do not say who paid for it. It is a small test, but I like that it shows where agents still need a person to check the work.
Microsoft opens a Work IQ preview for business data
This one is a little older. ERP Today reported on 30 September that Microsoft opened a public preview of Work IQ for Dynamics 365 and Power Platform data. It gives Microsoft's Copilot and agents a shared model of a company's business data. Rollout continues through October, per a Microsoft post.
It has three parts: a semantic model of the data, business skills (reusable company procedures), and governed actions. Governed actions are proposed changes that go through approvals and run within the user's permissions. Work IQ agents run under Microsoft Agent 365, which Microsoft says lets IT see how they use business data. Billing is usage-based through Copilot Credits.
The article warns that agent answers will depend partly on how well a company has kept its setup tidy. Dynamics 365 Finance and Operations coverage arrives in late October, so ERP scenarios are not part of the preview yet.
My take: Microsoft makes and sells all of this, and the article mostly relays Microsoft's own description. I read one source and saw no outside test. The approval step for actions is a good idea, but I would want to see it work with real company data.
What I take from this week
- Standards for agents are being written now, and the companies writing them are the ones who gain, so I will watch who joins.
- Agent startups are raising big money, but many of the usage numbers come from the companies themselves.
- A fair outside test says the best agents still stumble on judgment-heavy work, so keep a person checking the result.
I will keep opening the original pages and telling you whose claim each number is. If a story has only one source, I will say so.
Sources
- The Next Web: Sierra announces Personal Agent Protocol, an open standard for personal AI agents
- The Next Web: Nous Research raises $90M to take its Hermes AI agent to businesses
- Dealroom: Nous Research raises $90M at $1.5B valuation to take Hermes Agent to businesses
- Crypto Briefing: Epoch's Automation Reports find frontier AI still can't do Epoch's job
- AlphaSignal: Epoch's AI Benchmark Finds Claude Fable 5.1 Still Needs Human Review
- ERP Today: Microsoft Opens Work IQ Preview for Dynamics 365 and Power Platform Data