AI agents are AI tools that do a job step by step, instead of just answering one question. I opened every source below myself before writing this. When a number comes from the company selling the product, I say so.
The US Army is reported to be building self-directed hacking agents
The Intercept reported on 19 October (reporter Sam Biddle) that US Army Cyber Command is working on a programme codenamed ECHOTRIBBLE. It is an early effort to turn AI agents into tools that hack on their own. The report is based on documents sent to possible contractors.
According to those documents, the agents would move from device to device like a worm, including cloud systems and "air-gapped" systems, and then destroy themselves when finished. The software "must operate with minimal human oversight". A Q&A document says "neutralizing" a target could mean disabling it for a while, disabling it completely, or deleting it.
Three outside experts warned about the risks. David Krueger, from the University of Montreal, said: "We've seen AI systems knowingly attack systems that were not intended targets." William Marcellino of RAND said current public AI technology lacks the judgment that humans use to hold themselves back. The Army's cyber command did not answer The Intercept's questions, including how it would prevent accidental damage.
My take: This is one outlet's reporting from documents it obtained. I have not seen the documents, and the Army has not commented on the record. The program is described as early. Still, the key point is simple: an agent that acts with little human checking is a very different thing from a chatbot, especially when it can spread.
Disclosure: Anthropic makes Claude, and I use Claude to help run this blog. Read this section with that in mind.
OpenAI buys Orchestra Bio to bring agents into drug development
On 20 October OpenAI announced it had bought Orchestra Bio, a San Francisco startup whose AI agents help drug companies keep development programmes on schedule. The price was not disclosed.
Per Decrypt, Orchestra's agents flag problems on their own, such as a failed manufacturing batch. They work out which activities and deadlines the problem affects, gather information and compare possible responses. Scientists and programme leaders still make the final decision. The Decoder describes the same idea: tracking trials, flagging timeline impacts and suggesting next steps.
Orchestra's team and platform will be folded into Rosalind, OpenAI's set of AI models and tools for biologists and chemists. Decrypt says it is not clear when the agents will show up inside Rosalind. The Decoder notes OpenAI already works with Novo Nordisk and Moderna. Decrypt puts Orchestra's funding at about $20 million, based on Caplight data.
My take: The goal of faster drug development is OpenAI's stated aim, not a result anyone has shown yet. The 15-year and under-10% figures are background numbers, not results from this deal. What I find useful is the design: the agent does the tracking and comparing, and a person decides.
Motive launches Agents that act on fleet problems
On 19 October Motive, a company that sells software to trucking and other fleets, launched Agents. Motive describes them as AI that acts when something goes wrong, instead of only sending an alert. They are available now to Motive customers.
Examples from Motive: an agent can spot early signs of driver fatigue and have Motive's assistant, Atlas, check in with the driver. It can warn about severe weather and suggest another route. It can explain a fault code, open a maintenance work order and mark a vehicle out of service. It can also flag too much engine idling. Chief product officer Hemant Banavar said: "Our customers don't need another alert. They need action."
Customers can set agents up from guided workflows, templates or a library of Motive-built agents. Motive says that early next year customers will be able to describe an agent in plain language.
My take: Both figures come from Motive. The $46 billion is Motive's estimate of what organisations lose each year to preventable problems. The 42% comes from Motive's own ROI report, where people picked up to five concerns from a list of 12. FleetOwner, which also covered the launch, says the figures are Motive's own and not independently verified. Marking a truck out of service is a big action, so I would want to know how much a person can review first.
Notion rebuilds its agents to run longer for less
On 19 October Notion co-founder Akshay Kothari published a post saying Notion Agent and Custom Agents now run on new infrastructure. The headline change: an agent can work on a single task for more than 24 hours, up from five minutes.
Notion also says agents connect to hundreds of tools through MCPs and connectors, run up to 50% faster, and cost about 30% less for Custom Agent runs in its testing. A model router called Auto picks a model for each task. Notion says that across more than 20 common use cases, Auto gave similar results to Opus 5.5 and Astra at about 90% lower average cost. A new "Review for me" mode uses a separate model to check whether an action is safe before the agent does it.
On pricing, Notion says Business and Enterprise seats get 1,500 Standard credits and 500 Select credits each month. New customers get this from 19 October. Existing customers move on 2 November. Technobezz reported the same credit details.
My take: Every speed and cost number here is from Notion's own testing. Notion also makes the tool where I keep this blog's drafts, so I am not a neutral reader. The "Review for me" idea is worth watching, because a second model checking the first is one way to add a safety step.
Disclosure: Anthropic makes Claude, and I use Claude to help run this blog. Read this section with that in mind.
What I take from this week
- Agents are being given more freedom to act on their own, from fleet trucks to drug programmes, so the question of who checks them matters more every week.
- The designs I liked keep a person or a second check in the loop: scientists decide at Orchestra, and Notion adds a safety review. The Army story is the opposite example.
- Most numbers this week come from the companies themselves. Treat them as claims until someone else tests them.
I will keep reading the sources myself and flag who is speaking. If a story only had one source, I will say so, as I did with the Army report.
Sources
- The Intercept: The Army Wants to Sic AI Agents on Its Enemies. What Could Go Wrong?
- The Decoder: OpenAI acquires Orchestra Bio to build AI agents that speed up drug development
- Decrypt: OpenAI Buys Orchestra Bio to Put AI Agents in Charge of Drug Development Grunt Work
- Motive: Motive Launches Agents: AI that Acts Before Problems Start
- FleetOwner: Motive launches AI agents to reduce truck collisions, breakdowns and operating costs
- Notion: Agents, rebuilt to do more for less
- Technobezz: Notion AI Adds Two Credit Pools and Longer Agent Tasks