Coté

Coté

How to code agentic AI tools in Java with goblins

The goblins get into agentic AI.

The above video is exciting for me: it’s me relearning programming, playing D&D with the robot, and coming up with a new type of way I can help out at work. In this introductory video I go over the basics of making a tool (an “MCP Server”) for Claude.

This tool is a very simple oracle that will answer yes/no questions. Oracles are a core part of solo role playing and introduce unknown twists and turns, help you come up with adventures on the fly, and so forth. I’m not sure an AI like Claude needs an oracle, it might be good enough at picking random results that lead to different adventure paths. Or maybe it’s not! AI’s work by typing out the next word (yeah, nerds: token) that logically comes next, so maybe it is very much not good at random results!

Anyhow, this oracle is super basic, but it shows how to make these tools with Spring AI. Java is used by millions of developers, especially enterprise developers, and Spring is used by many (“most all”?) of them. Spring AI makes creating these agentic tools really easy. As you’ll see in the video, setting up Claude is more difficult than making the actual tool!

I’m going to put out more videos, making more complex (and useful) tools. I’ll also explore running this on your own machine - I’m about to try on a 10 hours flight so I hope I can figure it out.

You can see the basic tools I’m making for this video series in my EasyChatDM project and my “real” tools in the ChatDM project.

Relative to your interests

  • DATEV Accelerates Tax and Finance Software Innovations with VMware Tanzu Platform - “We recently patched application platforms that host 18,000 containers four times in one week. That wouldn’t have been possible without all the automation we have with Tanzu Platform.” // DATEV eG modernized its application development and deployment pipeline using VMware Tanzu solutions, enabling rapid development and deployment of tax, finance, and business administration applications.

  • Three Methods for Channeling Shadow IT’s Energy - “For CIOs, the question shouldn’t be about whether to eliminate shadow IT but how to harness its potential while mitigating its dangers.” // Or fine them with “what I like to call DARC (dangerous, awfully conceived, redundant, or costly) solutions, they should face financial consequences.”

  • On Da Vinci and Boredom - “More than anything, observe a brilliant person for whom both the prospect and opportunity of boredom led him to follow his curiousity into whatever intellectual avenues it wanted to pursue, and then turning his imagination into product manifest in text and on canvas.”

  • Running MCP-Based Agents (Clients & Servers) on AWS - Running MCP Servers on AWS, with Spring AI.

  • Researchers suggest OpenAI trained AI models on paywalled O’Reilly books - “GPT-4o [likely] recognizes, and so has prior knowledge of, many non-public O’Reilly books published prior to its training cutoff date,” wrote the co-authors.” // I mean, we know that the AI people are doing regulatory arbitrage, breaking the law to get competitive advantage.

  • Picking Apart An Ebullient GenAI Spending Forecast - “Lovelock says that companies will be looking for third party software providers to add AI features rather than try to do all of this GenAI creation and integration themselves, which reflects the attitude that most IBM i shops have had thus far. Straw polling we have seen shows that companies are waiting for either IBM or application software makers to give them the tools to have AI functions in their applications.” // And, a lot of the big-ass estimate is AI PCs, so you can probably throw that out.

  • JFrog Survey Surfaces Limited DevSecOps Gains - All that DevSecOps marketing from a couple years ago didn’t stick. // “A global survey of 1,402 application developers, cybersecurity and IT operations professionals finds 71% work for organizations that, despite any potential vulnerabilities, still allow developers to download packages directly from the internet.”

  • The Strategy Behind MCP - Good theory on the business strategy for MCP: it keeps LLMs at the center of value. // “Anthropic is not positioned as a consumer brand. Everything in their actions and intent thus far has been positioned towards enterprises or power users.” // “you need the tools to disintermediate themselves. That, at its core, is the goal of MCP and in my opinion a core strategy for Anthropic.”

Wastebook

  • “No one hate-listens to a podcast.” Here.

  • “Face with bags under eyes.” New emoji for recent parents.

  • Featurecide: the slow killing of a product’s soul in pursuit of every trend that moves the needle on engagement metrics, no matter how disconnected it is from the original mission.” Here.

  • “We’ll do four hours of shit-talking nonsense.” Noah.

Logoff

I composed and sent this in my phone. You can tell the Substack product managers were like “sure, let’s make it work on iPhones, but don’t spend a lot of time on it.” Which is fine, really. I imagine there aren’t many people who want to sit down to a breakfast of typing a newsletter on their phone.

@cote@hachyderm.io, @cote@cote.io, @cote, https://proven.lol/a60da7, @cote@social.lol