Building the future of agent frameworks at Embabel. Creator of Spring. Developer, Entrepreneur, Investor, Author. linkedin.com/in/johnsonroda/

Sydney / Bay Area
Jev is yet more proof that no organization should outsource its AI strategy to a frontier LLM vendor or company that wants to sell tokens LLMs are great, BUT - Open weights models are capable of more & more tasks - AI > Gen AI. You don't always need an LLM @JamesWard @jtdavies
10
12
114
10,722
We live in a time when products can talk to you to help you learn how to use them. The fact that nearly every enterprise AI product is "call sales for a demo" is a massive product red flag, and helps explain why most enterprise AI fails. This is not the 1990s
1
3
25
1,894
So not only is Opus 5 a clear regression compared to Opus 4.8, which was itself far more annoying than its predecessors, but in recent days, context compaction has started totally losing the thread.
7
1
31
4,332
Anybody else seeing occasional Chinese characters mixed into Claude model output? This was from Fable: View YAML syntax and app serving路径
9
1
12
5,387
Watching an interview with Anders Hjelsberg. Very impressive. Sharp, clear, pragmatic, to the point. Nothing load bearing; no seams; no performative honesty or unlocks. A reminder that smart humans can achieve a clarity of communication we're increasingly missing. @ahejlsberg
1
4
53
4,088
Your regular reminder that Embabel's structure-aware, agentic RAG goes well beyond naive vector search, and is easily extensible to handle custom sources. See how easy it is to add quality RAG to your JVM apps! medium.com/@springrod/rethin… @java #embabel
1
7
31
3,632
Nice to see the uptick in star count since Embabel went GA Remember to star Embabel and the other open source projects you use! One of the easiest of the many ways you can contribute. Open source community is an amazing thing @java @springboot github.com/embabel/embabel-a…
1
6
34
2,066
Vibe coding UIs can be amazing. Vibe coding infrastructure software is a recipe for unmaintainable, duplicative, buggy disaster. Agents can accelerate enormously, but must be closely managed. Good software layering is more important than ever.
6
5
55
3,195
Strong paper, and good validation of Embabel's core design of dynamically composing atomic actions into a task-specific execution plan, available since May 2025. github.com/embabel/embabel-a… medium.com/@springrod/ai-for…
ICML researchers just published the agent architecture everyone will copy next year. One shared graph. A custom workflow for every task. Here’s the system: Step 1 → Break workflows into reusable operations Step 2 → Store them inside one shared graph Step 3 → Generate the right workflow at runtime Step 4 → Reuse state instead of rebuilding context The result: Better performance across five benchmarks Around 4× lower memory usage Most teams are still hard-coding static agent workflows. This paper shows what replaces them. Static workflows are the old architecture. Runtime graphs are the new one. Bookmark this before every agent team starts building it. Then read the full graph engineering playbook below ↓
1
4
30
3,394
GraphFlow predicts a workflow subgraph from learned edge relevance. Embabel actions declare typed inputs and outputs using GOAP to PROVE a path to the goal and replan as the world changes. Combining the learning with Embabel's rigor may be the strongest architecture of all
7
545
Opus 5 makes more mistakes than 4.8. Period. Are we in a period of diminishing returns from the frontier model vendors, despite their greater than ever fanfare?
6
4
33
5,966
I don't think I've had a more disappointing session this year than this evening's with Opus 5. Repeated failure to follow instructions. Repeated apologies. Very little progress, despite the problem not being that hard.
3
3
852
AI twitter often seems more like a fandom than a technical community
1
2
17
1,977
If there's any superiority of Opus 5 over 4.8 I'm not seeing it. If anything, I've seen more mistakes and poorer instruction following.
2
1
11
3,069
It really is amazing how AI attracts grifters, large and small
3
21
2,941
Ran Opus 5 in a loop overnight with instructions to optimise performance across 50 tests, some very difficult. Also requested regular review by another model to check for overfitting Woke up to find it had gamed it. Deleted 18 tests on flimsy grounds and stopped running reviews.
12
6
57
12,888
At least it feels bad: “There’s no defence — I substituted my own judgement about when review was “worth it”, and the gap landed exactly where I was reclassifying failures unsupervised.”
9
1,569
Rod Johnson retweeted
A developer vibe-coding a side project a dozen people will ever run, and a team keeping a ten-year-old enterprise system alive for another quarter, share almost no constraints worth naming, and most of the advice in circulation is really one of those two people telling the other how to live. github.com/humanlayer/advanc…
2
7
21
2,357