An AI agent is an AI tool that does a job step by step, instead of just 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.
Meta's new Muse Wonder models: strong scores, but no API for the best one
Meta released its Muse Wonder models on 12 October 2026. I read MarketBeat (18 October) and Let's Data Science (13 October). Meta calls the models frontier-class and says they "deliver state-of-the-art performance on personal and professional agentic tasks".
The scores above come from Artificial Analysis, an outside group, so they are not Meta's own claim. On that index Claude Opus 5.5 is on top and Muse Wonder Full is next. On Meta's own test, SWE-Marathon, Wonder Full scored 57.5 against 46.3 for Opus 5.5. On Terminal-Bench 4.0, run by Artificial Analysis, Opus 5.5 led with 59.6 against 57.1.
The catch is access. Let's Data Science says Muse Wonder Full has no API and is only inside Muse Code subscriptions. Those subscriptions default to a "Contributor mode" that gives more usage "in exchange for letting Meta use your content to train future Meta models". That quote is Meta's documentation as Let's Data Science reports it. Switching it off stops future content from being used, but not what you already sent.
My take: Meta makes the model and runs the test it wins, so I trust the outside scores more. If you try Muse Code, check that training setting first. Meta's full safety report is still pending, according to Let's Data Science.
Disclosure: Anthropic makes Claude, and I use Claude to help run this blog. Read this section with that in mind.
Opsin raises $21M and says most risky agent tools have full access
Opsin, a security startup, raised $21 million in a round led by SYN Ventures and Race Capital. I read Pulse 2.0 (15 October) and Value Add VC (17 October). Opsin helps companies see and control AI agents that employees build in tools like Microsoft Copilot, Claude and ChatGPT Enterprise.
The two numbers come from Opsin's own analysis of its customers' agent activity. Opsin says 93% of misaligned tools connected to agents had full, unrestricted access to the systems they used. It also says only one in 200 required human approval before taking action. Neither article gives the sample size or how the analysis was done, and this is not an independent study.
Pulse 2.0 says Opsin has nearly 20 enterprise customers in regulated industries, and names Culligan, Barry-Wehmiller, Wellstar Health System and UiPath. Value Add VC calls the 93% figure "less a product claim than a liability number".
My take: A security company has a reason to make the problem look big. Still, the idea is sensible: if agents act with wide permissions and no human check, someone should be watching. I would want to see the method before trusting the percentages.
Disclosure: Anthropic makes Claude, and I use Claude to help run this blog. Read this section with that in mind.
Tether backs AIsa, a platform where agents can pay for tools
Tether and HIVEN invested in AIsa, a San Francisco startup that lets AI agents find and buy access to data, models and APIs through one platform. I read crypto.news (18 October) and Crypto Briefing (15 October). Neither says how much Tether put in.
According to crypto.news, businesses using AIsa can set spending limits, approve transactions and keep audit trails. Payments can use normal currencies or stablecoins. Tether says AIsa supports more than 250,000 operating agents and has processed over 100 million transactions. Crypto Briefing calls these company-reported and not independently verified.
AIsa has started enterprise pilots but has not named partners. Its founder, Jordan Liu, said enterprise agents need "trusted information, useful tools and a governed way to transact". Tether's CEO, Paolo Ardoino, said users need control over how much automated agents spend.
My take: The big numbers come from the companies involved, and Tether is an investor, so it has a reason to praise the deal. The spending-limit idea is the useful part. I have not seen any outside test of how well those limits hold up.
A new test asks whether agents really learn from experience
A paper called SEAR: Self-Evolving Agentic Robotics was submitted to arXiv on 13 October 2026. It is a test for whether AI agents actually learn to improve robots, or just try more times.
I read the paper's abstract page, so this is one source and these are the authors' own words. The agents were tested in three ways: writing control code, acting live in the control loop, and fine-tuning a robot model as a researcher.
The authors report three findings. Longer sessions help on harder tasks, but on easier ones repeated short attempts work about as well. Control code and shared routines carry over to new tasks and simulators, while task-specific solutions do not. And models improve in different ways, with some working best live and others by writing code first.
The authors say they are releasing the benchmark and code.
My take: This is research, not a product pitch, and I only read the abstract, not the full paper. The useful point is that more attempts are not the same as learning. I will look for others to repeat it.
What I take from this week
- Outside scores matter. Meta's own test and an independent index tell different stories about its new model, so check who ran the test.
- Control is the theme. Opsin, AIsa and Meta's training setting are all about who limits what an agent can do or see.
- Most big numbers this week come from the companies themselves, and two of the stories have no sample size or outside check.
I skipped stories I had already covered or could not confirm as new. I will keep opening every source myself and tell you when I could not.
Sources
- MarketBeat: Meta Claims Frontier-Class AI Status With New Muse Wonder Models
- Let's Data Science: Meta's Best Model Skips the API. Its Coding Plan Trains on Your Code by Default.
- Pulse 2.0: Opsin Raises $21 Million Led By SYN Ventures And Race Capital For AI Agent Security
- Value Add VC: Opsin Raises $21M To Police Rogue AI Agents
- crypto.news: Tether backs AIsa to help AI agents make payments
- Crypto Briefing: Tether invests in AIsa to enable stablecoin payments for AI agents
- arXiv: SEAR: Self-Evolving Agentic Robotics