Agentic-skiils for ERDDAP™ redux

[Roy Mendelssohn - NOAA Federal]

Agentic skills (https://agentskills.io/home) are "a lightweight, open format for extending AI agent capabilities with specialized knowledge and workflows.”. They can provide many of the same benefits as a Model Context Protocol (MCP) server, but are markdown files, not code. Because skills are markdown files, they are easily edited and can be customized. An MCP server has the advantage that it can execute code if you have an agent that can not execute code, but for a variety of reasons touched on below it may well be preferable to have agents generate but not execute code. But as usual in these cases it is not either/or, and this is just another tool to make ERDDAP™ work better with agents.

I have been working on agentic-skills for ERDDAP™. I opened a discussion at https://github.com/ERDDAP/erddap/discussions/511, the actual skills, still being refined, are at https://github.com/rmendels/erddap_skills. Perhaps of most interest to ERDDAP™ admins there is now a skill for GenerateDatasetsXml at https://github.com/rmendels/GenerateDatasetsXml-skill, which is still being refined, though in my tests works pretty well. As mentioned above, one advantage of agentic-skills is they can readily be customized, so for example you can customize the GenerateDatasetsXml skill such that certain tags with given values are always added, or that certain tags always have given values.

Some things I have learned using agents-skills, perhaps well known but indulge an old guy:

1. Keep chats short. Every time you add to a chat and hit send, unless there has been a big improvement in agents, requires the entire chat to be read in again by the agent, which burns through your tokens. If possible save intermediate steps and start a new chat.

2. Though not necessary, ask the agent to review the relevant skills before starting. This tends to produce better results.

3. Most skills have certain keywords, the more you use those in the prompt the more the agent knows to use the skill. In general, the more information relevant to the skill you can provide the agent the better the result.

4. Insist that intermediate results be shown and approved. I have seen some really wild or overly complex responses generated by agents (though I have also seen some really good responses). This is a good way to keep the responses focused. Of course the agent providers tell you to set a goal and let the agent rip, because that will eat up all of your tokens fast, and it also puts a lot of trust in the agent.

5. Ask for code as the output, do not have the agent run the code. This provides extra security, as well as providing reproducible results. An MCP server that runs code and just gives the results may not provide reproducible content.

I have also started to use the “Open Knowledge Format (OKF)” (https://github.com/GoogleCloudPlatform/knowledge-catalog/tree/main/okf) to keep a history of what I have “learned” in particular with respect to agents, though it is not limited to agents. Many agents can be set up to work with OKF files, usually after a session with an agent I ask it to update my OKF files with what has been learned.

If you use any of the skills and have any comments or suggestions for improvements or additions, either add to the discussion on the ERDDAP™ Github page, add a discussion to my Github pages linked above, or email me directly.

Thanks,

-Roy