I can't express with enough words how happy I'm to join the amazing group of @Java_Champions!
I need to say thank you to the global @java community for being so good to me since the day one.
Thank you also to my fellows from @soujava: it wouldn't be possible without you!
#java
40% of AI agent projects will be canceled by 2027. That's not me, it's Gartner. And the reason has a name: agent washing.
They did a survey and found that a lot of what people called agent projects were chatbot projects, RPA, or assistants, even if built on LLMs, purely assistants.
What was missing for them to count as agents? Being able to take some kind of action autonomously. A chatbot doesn't. An RPA doesn't. Autonomy doesn't mean doing things left and right, but some level of autonomy. Without it, it's not an agent. And plenty of people are selling agent projects without building agents.
There's another number, from an MIT report published in 2025: only 5% of AI tools built inside companies reach production. Even worse. 95% of AI projects never get there. The same report talks about 30 to 40 billion dollars invested in generative AI, with 95% of companies getting zero return.
I want to look at that 5%. While 95% goes down the drain, there's demand that is reaching production, generating value, built for results, built looking at the client, the company, the project.
If you really want to work with AI, where you are or at another company, you need to aim at those companies, those projects, those cases that actually reach production.
A software developer exists to solve problems. That's our career. And we only solve problems if what we build goes to production. What only runs on your machine maybe solves your own problem. Is someone paying you to solve that? If yes, great. Most of the time, what's only on our machine helps nobody else.
So aim for the 5%: what goes to production, solves someone's problem, makes someone more productive, gives someone a less miserable day. Your next project or your next company may be in there.
The AI conversation went from "AI will take our jobs, AI will replace me" to "AI will kill me. AI will kill everyone, humanity will go extinct because of AI."
And many people ask:
- "What do I do if AI replaces me? What do I do if AI kills me?"
Whatever I think about either case, something I kept thinking about is that part of this urge for an answer in extreme cases is that we always want someone to give us the answer. And that's fine. We look for answers in others.
In these extreme cases the doubt comes from very divergent opinions from very relevant people, sometimes relevant in different areas, which may be why the opinions differ. One says AI won't do anything, the other says AI will kill everyone. One says everyone will work with AI, the other says nobody will work anymore. Relevant people on both sides, and we're left in doubt.
My point: we're adults. I assume you are. We need to be able to make our own decisions and take our own positions. Even if your decision is to follow someone else's. That's fine.
But you can't depend on someone on the internet telling you what to do, what's right and wrong. You need to make that judgment as an adult, because in the end, even if you follow someone's judgment, the consequences fall on you.
You can try to blame someone. Say AI took everyone's job:
- "I should've done it differently, but I followed what so-and-so told me."
OK, but you followed it. So-and-so just gave an opinion, with good intentions or not.
I keep telling software engineers that even if AI writes your code, you're responsible for it. In everything in this story, the responsibility is mine and yours.
So listen to whoever you need to, study what you need to, embrace whatever conspiracy theories you want, and in the end own the decision.
AI made our work as developers much easier. Faster code, faster troubleshooting, faster debugging, faster exploring a new codebase.
But it created a kind of problem we didn't have, at least not at this scale: silent errors.
When you compile an application and there's a compilation error, it's in your face. No doubt. You run a build and it breaks. No doubt either.
When an LLM gets it wrong, most of the time it doesn't throw an error in your face. It hallucinates.
A colleague of mine wrote in a post that he was working with an LLM and it invented a Quarkus extension that doesn't exist. It thought it was reasonable. Plausible. That's the plausibility principle.
When the hallucination comes in a direct answer, you glance at it and notice. You deal with it. The problem is when it happens at the reasoning level and never reaches you directly.
One context engineering concept for this is context poisoning. A hallucination becomes part of the context and the agent starts deciding based on it. A close cousin is when it takes a wrong direction because of a hallucination and keeps going that way.
The answer: pay attention to your agent's context. People kind of stopped talking about context engineering after a hype earlier in 2026. The term came up around mid 2025, and it took off with personal agents like OpenClaw, Hermes, now Grok Bot. The better the context, the better they work and the less they get wrong.
That's true for your code too. A lot of people think:
- "The agent writes the code now, I don't need to care about clean code, project organization, the domain."
It's the opposite. If your code is a mess, it's a mess for the agent. It may take longer than you to get lost, but it will.
So do what a compiler does: put restrictions in place so that when something goes wrong, it stops and tells you.
You're probably asking AI to do things you shouldn't be using AI for.
The AIs we use today are generative AI. AI has existed for at least 50 years, but with generative AI, what you expect is that it generates something, creates something.
Whenever you send AI something repeatable, predictable, you almost certainly shouldn't be using AI for it.
A silly example: ask it for one plus one. When it gives you two, in most cases it didn't do the math. It looked at its training and saw that probabilistically one plus one should be two. That's why some AIs have specific modes for math, a different area of training. When you see AI solving old complex math problems, it's because they're putting math training into models. For basic calculations, it doesn't do the math.
Bringing this to software development: a lot of what people are automating with AI should be a script, a routine, a batch, a cron, something that runs every day and executes that predictable thing.
Then someone says:
- "But, Elder, I want AI to look at my issues. A user opens an issue on my app, I want AI to understand it, write the plan, maybe the implementation, open the pull request. I can't?"
Calm down.
Understanding the issue, writing the plan and writing the fix, yes, AI does that.
What you shouldn't use AI for is going through your issues every day. That's predictable. Same repository, always issues. Write a script, run it on a cron. When it finds an issue, then you connect it to your AI, which does what isn't the same for every issue. Each issue is different, each ticket, each client.
Same with customer service. Use AI to answer the client. But receiving the WhatsApp message and triaging it is deterministic. Same WhatsApp, same webhook, same API.
Your AI usage gets more consistent, more stable, and much cheaper, because you're probably burning a ton of tokens making AI do what it wasn't made for.
A lot of people are still holding two, three jobs. I've seen people with four. It took off during the pandemic, when remote work was unavoidable.
Some companies found out and fired people. Some put it in the contract. Some don't care: work 15 jobs, as long as you deliver.
I had someone in my mentorship with three jobs. A very specific situation, I totally understood. But I told him what I'll tell you.
If that's your case now, each one knows their own challenge, but do everything you can to make it temporary. It isn't sustainable, especially if you take each job seriously. Some people have three jobs and don't care about any of them. This isn't for them.
Treat it as:
- "I'm doing this now so that soon I won't need three. I'll need two, then one."
Otherwise you'll wear yourself out.
How? Take from everything you do the thing you do best and develop that profile, so you can earn in one job what you earn in three. That's human. That's manageable in the long run. I'm getting close to 30 years of career. I know people with 40, 50. It's long. You don't want to spend it doing three jobs.
One last thing. I've heard:
- "I only feel safe with at least three. If I'm in one and get fired, I'm at their mercy."
I get it. But people stacking separate contracts never get a really good position. A really good position will probably require more than two or three hours a day. You'll have to dedicate yourself.
So be careful you're not depriving yourself of responsibilities that would take you to another level in the name of a supposed security that's really a fear of taking on a challenge where you'd be well paid to do one job.
15-year anniversary! I have attended every JavaOne Conference since 2012 *
I am super excited to be on stage with the Brazilian Java Man Bruno Souza and talk about the future of the Java Community!
*(there was a small gap during the pandemic)
we don't have to live this way. senior engineers have the duty of building guardrails and teaching the next generation of programmers the practices that lead to high quality software.
AIs may write most of the code, but you still decide what ships. we should never lose our taste for quality software.
i made this analogy a while back: everyone can cook at home but yet restaurants still exist. the floor is rising, but so is the ceiling. the best software will always be made by those who care.
Lately I’ve been feeling depressed because decades of our hard work is completely gone like it never existed. I heard from others that they also find it very hard to dial into the new norm of low quality software engineering.
Argo Rollouts can roll back failed deployments. Can AI agents help fix them on the fly? 🔧
At #Devnexus2027, @kevindubois from @IBM explores Java-based agents that analyse failures, fix code or manifests, create PRs & notify teams.
🔗 devnexus.com/events/fix-prod…
If your only networking tool is social media, you're counting on rented land.
This feed you're reading, as free as it is, as long as you've used it, isn't yours. You can lose it any time. Accounts get hacked. Someone files a crazy report and you lose the account. It's not rare.
A lot of people rely only on that to stay in touch with people in their market, people they've worked with. Out of sight, out of mind, so we try to be seen to be remembered.
It's fragile. Not just because you can lose the account. Purely digital contact is great for scale, but it's still very shallow.
Notice how people leave an in-person event, a conference, a meetup, a user group, posting photos:
- "It was amazing, I met so and so."
They leave with a different energy. Meeting in person is different.
I've worked with the same team for about six, seven years. During the pandemic it was all online for two, three years. The first time we met in person was at an event in Atlanta. A big table, everyone speaking English because of all the nationalities. People started leaving, until it was just me and my colleague Alex Soto, who's Spanish. So I kept the conversation going in Spanish. He looked at me and said:
- "Wait, you speak Spanish?"
I said:
- "A little."
And he said:
- "I've been talking to you for two years and didn't know."
From there came other work, we started helping each other across Latin America. If we hadn't met in person, maybe he'd still think I only speak English.
I prefer remote work a thousand times. But some opportunities, insights and conversations only happen face to face.
So don't hand your whole professional network to a social network.
In tech we've always dealt with knowledge obsolescence. You learn a tool and soon there's a better one, or a better version. You learn a framework, it breaks the contract, you update, a feature appears, one you used goes away.
With AI that went exponential.
A while back I was telling my students:
- "What I planned for the Java AI Specialist program in February, I've already had to update."
Some things I hadn't recorded yet, I changed completely. Some things I already delivered, when I finish the course I'll go back and adjust. Things are changing that fast.
"What do I study, where do I invest my time" was always in a dev's head. Now I imagine much more, and some people must be even more anxious.
So maybe the question now is:
- "So how do I stay up to date now that it's even faster?"
The answer is the same. Focus on fundamentals. On what doesn't change, or changes very little.
Distributed systems: the theory behind it is the same since the 80s. Tools changed, infrastructure changed. The classic books are still very valid. The Saga pattern, for distributed transactions, came from the 80s, was revived with microservices, and is being revived again with AI agents. Same fundamental, different tool.
When you learn to separate the fundamental from the tool, even minimally, you get an advantage for your whole career. Tools change, and you learn the new ones very fast because you know what's behind them. Your learning becomes syntax, configuration, the vocabulary of that new thing implementing something you already knew.
No matter how fast the tools evolve, the fundamentals stay. Dedicate even more time to them.
Every March and every September we get a new Java release. Since Java 10 it's every six months.
And every time, developers react in a few ways.
- "Oh no, I'm still on Java 8 and now there's 25, 26, 27. One more version I'm behind."
- "I'm going to read the whole release, dissect it, understand everything."
- "Too big. I haven't even read the one from six months ago."
All of them have a point. Mostly bad ones.
What I do, for a long time now, is have a strategy to read the release. It improved my practical results and my head. I don't get stressed anymore.
First, I glance at all the items.
Second, I classify them: does this affect me as a developer or not? If it changes how I write code, a new feature, a removed one, a changed one, that interests me. If it changes some internal cycle of the JVM, it's not that I don't care, it just doesn't affect me directly. I'll write the same code tomorrow. Indirect goes down the priority list.
Third, among the direct ones: is it final, or still preview or incubator? Structured Concurrency, for example, is still changing. I built Structured Concurrency demos I had to rewrite. That was expected. If it's still changing, it gets lower priority, because anything I build on it may change.
Out of a release with around 15 new items, what you really need to study is five, six, seven. Rarely ten.
You've limited your scope a lot, and now you know what you read really matters to you, and matters now. That's how I build my "Java 25, what really matters" and "Java 26, what really matters" talks.
When we talk about AI for software development, most of the market, maybe 80% if we use a Pareto, is thinking about AI-assisted development. Claude, Codex, someone to write my code, test it, review it, deploy it, review my pull requests, open issues. That's great. I use it, I love it.
What I see most professionals missing right now is using AI in your software. Connecting the application you're building to an AI so the AI is part of the software.
The basic example is customer support. Put a chatbot on WhatsApp that talks to your customer. Cool. I'm talking about more.
Say you have an ERP in your company and an AI connected to it handling the approval flow. There's a workflow, several approvers, and each one has to look at what's being approved, the subject, the implications, and give an opinion. You could have an agent looking at all those variables and handing the approver a recommendation:
- "Here's the case, the client, the money, the implications. I recommend this because of this and that."
The approver already has context, something pre-chewed, but the decision is still theirs. Human in the loop.
Very few people are doing this. And for me it's by far where we can gain the most with AI right now, because these systems have been in production for decades. Take a project at your company that has worked for a long time and add an AI leg inside it. In data analysis, for example.
You might ask:
- "What if it gets it wrong?"
It will. Not if, when. It's not there to do your job. It's there to give you more input, more insight, maybe improve your customer's experience.
Today you probably have an automated email reply where everyone gets the same answer. Use AI to customize it so it feels closer to that customer.
Recently a student of mine in the Java AI Specialist program posted on our Discord that he'd built a POC with AI for a project at his company. He left that meeting already invited to join an AI project there.
What he did should be obvious to all of us. Unfortunately, it isn't.
When someone wants to work with a technology, what do most people do? They ask:
- "Give me a chance, I'm here, I know the business."
That's a possible approach.
What he did, and we'd talked about it in a mentorship session: he built a POC that minimally solved a problem he knew the company had, and showed it. What stayed in the head of the person who says yes or no was:
- "Without me asking, this guy already solved this problem at a small scale. So yes, I'll give him the chance to solve it at company scale."
That goes for your whole career.
Want a promotion? Most people go:
- "I'd like a promotion."
The person who approves it asks:
- "Why?"
And they answer:
- "Because I've been here a long time. Because I do my job right."
Doing your job and being there while you're paid is the minimum.
Or you go:
- "In the last year I had these deliveries, they solved these problems, one brought more money, one cut costs, one improved the customer experience."
You're showing you contribute to the business. The promotion may still not come, but the approach is completely different.
Same if you want to upgrade Java:
- "This version has records, virtual threads, scoped values."
The person approving often doesn't know what that is, or it makes no difference to them. Now if you say:
- "This solves a problem of yours."
They'll at least hear you.
Works for marriage too. Want to convince your spouse? Bring an argument that makes sense to them.