Hermes, an open-source AI agent from Nous Research, is being added to Microsoft Teams, where employees can hand it work like financial analysis, a document draft or sorting an inbox. For a team, the interesting part is a feature called Collective Wisdom, which lets one person's best way of doing a task become everyone's starting point.
The agent writes the skill, so nobody has to
Hermes already did half of this for individuals. After a complex task goes well, it saves the method as a skill: a short written description plus code that runs the stable steps the same way each time. Sharing skills is not new either, since Claude has let company administrators hand them out to staff since December 2025. But those skills are things a person has to sit down and build.
Collective Wisdom adds the step between. An employee asks Hermes which of their skills are worth sharing, proposes them, and after a review they are published for the team. Nobody writes a procedure document. The procedure is a by-product of work that was already done.
What it looks like in a finance team
Picture a finance team that closes its books every month. One analyst has learned to check the freight charges against the carrier's invoices before posting, because that is where the errors hide. Without an agent, that habit lives in her head, and a new hire learns it in the third month, after a mistake.
With Hermes, her agent saves the routine after a close that went smoothly. She proposes it, the controller reads it and approves it, and the new hire's agent follows it on the first close. If the routine saves each of six people half a day a month, the team gets three working days back every month, and the new hire never makes the mistake.
The same pattern fits a shop owner's supplier reorders, a recruiter's first screening of applications, or a school department's term reports. Anywhere one person has found a better way and the others have not noticed, the better way can now travel on its own.
The review is now the job
A skill runs the same way every time, so a good one spreads fast and so does a bad one. A shortcut that worked in one lucky month can become the whole team's habit within days. The person who approves skills is doing what the slow process of colleagues learning from each other used to do: deciding which way is right.
That also turns the 1990s failure around. The old systems drowned in material nobody had time to search; the risk now is a short library that everyone trusts too much. Nous Research promises that the library stays the company's own and is not training data for an AI vendor, which makes it an asset worth looking after.
A manager can test a skill with one question before approving it: would I be happy if everyone on the team did the task exactly this way from tomorrow? If the answer is "mostly", edit it first. Knowledge used to leave the building when a good employee did. Now it can stay, as long as someone reads it before it ships.