My conference participation has dropped sharply since I left DevRel for a regular office engineering job. In 2023, I was proud to have had 70 talks; I’ll finish this year with 4, including one online and one quickie. It has one beneficial side effect, though. Because I go to so few, I take the time to attend. Last Tuesday, I was lucky to attend and talk at the Porto TechHub conference, for the 4th year.
Here are the talks I attended and their respective summary. If you attended a talk or are the presenter, feel free to correct my possible misunderstandings.
Stop Prompting, Start Planning: Reliable Agentic AI with Embabel by Patrick Baumgartner
I’m cheating with this one: I couldn’t attend it, as it was running in parallel with one of the other talks. However, Patrick presented it at the Geneva JUG last month.
Patrick introduced the subject by mentioning the core issue with LLMs: their non-deterministic nature. Two guardrails can help with that.
- The first is Goal-Oriented Action Planning. GOAP is a planner that lets an agent decide its course of action by providing it with a context, a goal, and a set of actions.
- The second is standard programmatic constructs, e.g., conditionals (
if).
Patrick then described Embabel, which aims to leverage both:
Embabel is an Agentic AI framework for the JVM. Python has long been the go-to for machine learning experiments, thanks to its ease, ecosystem richness, and data scientist-friendly tools. However, it often struggles when you try to move from experiment to real-world, production-scale AI.
Embabel relies on Java code, types, annotations, and Spring so you can program steps in a workflow.
The wiring from one step to the next is inferred via types.
If the method taskA() returns TypeX, and the method taskB() accepts TypeX as a parameter, then A and B are steps, and B follows A.
Patrick showed several demos grounded in business use cases and the relevant code to highlight his points. At the end, he also did a demo of a visualization tool for the code.
|
My opinion
I like Spring and strong typing, even though I could nitpick about having a proper DSL instead of annotations. In any case, I must admit that I fail to see the value proposition in Embabel. We have plenty of workflow engines on the JVM. I started my Java career on one. Today, Camunda offers a modern Open Source platform fully compatible with BPMN. BPMN has the benefit of being backed by a specification. Icing on the cake, many software components provide visualization of any BPMN-based workflow. One could achieve the same benefits as Embabel with a workflow task calling an LLM. |
The Agentic Loops Every Developer Should Know by Rui Quintino
I expected a talk in line with the title and the description. I missed the first 5 minutes, but the rest was mainly demos of what the speaker built with coding assistants. For example, he described how he wanted to edit a Word document in parallel with his assistant. However, Microsoft Word sets a lock so you can’t do it. Thus, he used Claude Code to write his own Word clone to allow it.
|
My opinion
The speaker was very enthusiastic about what he achieved, and I can’t deny that it was great, but I didn’t get anything actionable out of the talk. |
Comparing Agentic AI Frameworks for Java by Sandra Ahlgrimm and Timo Salm
The talk compares Spring AI, LangChain4j, and Embabel through the lens of an AI nutrition planner.
| Capability | Spring AI | LangChain4j | Embabel |
|---|---|---|---|
Deterministic workflows |
Advisors & plain Java code |
Workflow agents |
@State and planner |
Pure agents |
Tool calling loop |
Supervisor agent |
Built in |
Memory and state |
ChatMemory advisors |
AgenticScope |
Blackboard |
Agentic patterns |
Experimental |
Built in, some experimental |
Built in, some experimental |
Observability |
Model level only |
Agent listeners |
Agent level |
Testing |
Mocks, evaluators |
Mostly manual |
Advanced testing support |
MCP |
Client and server |
No HTTP server support |
Client and server |
A2A |
Experimental |
Client only |
Experimental |
Maturity |
|
Agentic experimental |
GA, plus some experimental features |
Legend: Built-in | Experimental or some effort | Limited |
|||
|
My opinion
I found the talk to be to the point. The table above summarizes the comparison. I’m happy to have it if I need help choosing in the next 6 months. I also liked that the two speakers work for competing companies, Broadcom and Microsoft. Timo, from Broadcom, didn’t push Spring AI over the other two. Sandra, from Microsoft, is a committer on LangChain4j, which Microsoft doesn’t own. |
Material:
What’s New in Git 3 by Rob Richardson
The talk focuses on the 4 features that become the defaults in Git 3:
- The hash moves from SHA-1 to SHA-256
- Reference storage moves from loose files to reftable
- The first branch moves from
mastertomain - Bare repositories are now
safeby default
The speaker explained each item, its current state, and its future state. He also reiterated several times during the talk: these are the new defaults of existing options. Using Git 3 on a repo created with Git 2 defaults will work flawlessly. He also demoed how to migrate such a repo to the new defaults, one item at a time.
According to Rob, the biggest obstacle to releasing Git 3 is the lack of support from Git hosting providers.
|
My opinion
I wanted a non-AI talk, and I wasn’t disappointed. I particularly liked the fact that it was targeted at Git users, not Git internals experts. I’m now able to understand each item and how its default changes between Git 2 and 3. |
Conclusion
Porto TechHub was a very enjoyable experience. I’ve been a regular speaker, and if the chance presents itself, I’ll be happy to come again. Many thanks to the team for the flawless speaker experience.