SK Square has attached about 90 AI agents to its investment operations. They are said to be used for scanning the market and finding companies worth investing in. Depending on the material, the number is written as "90" or as "some 90," so the line between what is in operation and what has merely been built is blurred. The company also set up an in-house knowledge hub along with the agents.
Once the scale grows this large, talk of performance moves to the back.
Agents call tools on their own, open data and send things outside. It means that once a person gives an instruction, the agent takes several steps by itself after that. When it works well it is convenient, but when it works wrongly the error spreads across several steps.
Hana Card is taking part in the government's "AI for All (모두의 AI)" project to develop "AI Agent Pay." The service is aimed at situations where an agent makes payments in a person's place. Ant Group, Visa and Mastercard are each pushing their own agent payment standards. A figure of 7,000 trillion won is cited for the size of the related market, but it does not come with the year the forecast is based on, so it is hard to take down as is.
Payment is an act that is hard to reverse. So how the delegation limits and the points of user reconfirmation are designed decides whether the service succeeds or fails. That design is a part that will become visible as Hana Card's project proceeds.
There has been movement on the technology side as well. A control method was introduced that inspects the commands an agent issues just before execution and filters out actions that delete data carelessly or take it outside. Nothing was given about who supplies it or how well it blocks.
halo-record is published as version 0.2.42 as of September 9, and it leaves one record each for an agent's tool calls, model calls, data access and approval actions. The records are joined to one another with a hash chain. If one is altered after the fact, the chain no longer matches and it shows. The developers say it runs without outside libraries.
AgentTrust is a free scanner that goes over MCP servers and grades them from A to F, and it had graded 50 servers as of its release. MCP refers to the connection specification used when an AI model accesses outside tools or data. It produces inspection results matched to OWASP criteria, and it also includes a function that blocks deployment when a problem is found during the build process.
Running it on its own package returned a score of 90, grade A, with zero issues flagged. For both tools, no record of third-party verification has been given.
A company that adopted agents and a company that released tools to block agents appeared in the same week. It means they have passed the demonstration stage and entered real work.
Whether a company has written down what an agent that has entered its operations must not do cannot be known from outside. The difference between a company with that document and one without it stays invisible in ordinary times and shows up all at once when an accident happens. Whether the list of companies an agent produced can be traced line by line through the data it went through depends on whether it was designed to leave those records.
