Series
Learning Embabel by building with it, in order — planning, testing without a model, tools, human-in-the-loop, tracing, exposing it over MCP.
Part 1
My first afternoon with Embabel, the JVM agent framework. I added one action too many, the planner quietly stopped early, and a cryptic get(...) must not be null taught me how GOAP, goals, and AgentInvocation actually fit together.
Part 2
The non-deterministic part of an agent is the LLM. Box it out and the rest is ordinary, deterministic code you can test. Here's how I test Embabel agents on the JVM without spending a single token.
Part 3
An LLM cannot know a live account balance, and in banking a confidently hallucinated number is a disaster. The fix is a tool. Here's how @LlmTool works in Embabel on the JVM, plus the one API gotcha that cost me a compile.
Part 4
In regulated work the dangerous agent is the one that acts without asking. Here's how I made an Embabel agent pause for human approval before moving money -- with the approval built into the type graph, not an if-check.
Part 5
I pointed my Embabel agents at a local Grafana LGTM stack and saw nothing. Three configuration gotchas later, every GOAP planning step, LLM call and tool loop shows up as a span in Tempo. Here is the full walkthrough, with screenshots.
Part 6
Turning my JVM agents into MCP tools took one dependency and one annotation. It also took three silent failures -- a starter that suppresses the web server, a shell that eats your CLI flags, and an annotation default that publishes exactly zero tools without telling you.
Part 7
My agent refused to run a single action. The @Condition gate was FALSE and nothing in the agent could ever make it TRUE, because Embabel builds an action's effects from what you declare, not from what your condition reads.