Will developers be jobless? Yes. But not all of them
The talk’s Czech title was “A vývojáři nebudou mít co žrát”: devs will have nothing to eat. The translation is polite. The original is not.
The spoiler
Will developers be jobless? Yes. But not all of them. That’s the whole talk; the rest is about which half of the room you’re in.
Where we actually are
Six months ago, this was “the future”. Today you give an agent a task, not a prompt. It plans, writes, tests and opens the merge request. It runs unattended, on your own hardware or someone’s cloud. It wakes itself up, and asks you when it’s stuck. You’re no longer prompting. You’re delegating.
It’s real and it’s plural: Devin, Codex, Jules, Copilot Agent, Factory, Replit, OpenClaw, and a new one every week. The dates make the point better than the logos:
NOV 2022 ChatGPT: public, free, and very stupid NOV 2025 OpenClaw: one developer, a side project MAR 2026 Replit Agent 4: parallel task forking APR 2026 Factory raises a $150M Series C 2026 Devin · Codex · Jules · Copilot Agent ship PRs
Three years from a chatbot that couldn’t do arithmetic to “it opened the merge request”.
My opinion: engineering as we knew it has ended
This is opinion, not data. Every generation thinks its change is the big one, and I might be wrong too. But the shape is the same as before at a very different speed: roughly a hundred years for the industrial revolution, forty for electricity, ten for the dot-com era, five for mobile, and eighteen months for this. It’s also the first change where the thing being automated is the automating itself.
Who doesn’t adapt, dies. That’s why I’m not afraid. And honestly, I have no idea where this leads. Maybe trouble, maybe not, maybe it pops. Self-hosted models, hardware and energy are a different talk.
So we built another one
Yes, the world needed one more of these. It started in June 2026 as Cadence, a local, single-user experiment: capture tasks all day, press play, and it plans, implements, verifies and delivers. It worked. That was the problem: it stopped being a toy and became something we had to build properly, for the whole company.
The flow is modelled on a real team. Every stage is a role, not a prompt, and a return is normal, not a failure:
TICKET
↓
REFINE ──→ ARCHITECT ──→ IMPLEMENT ──→ CI
↑ ↑ ↑ │
└─────────────┴─── return ────┴───────────┤
↓
RETRO ←── MERGE ←── HUMAN ←── REVIEW ←┘
↓
KNOWLEDGE BASE ─→ every future task
The reviewer never sees the implementer’s session: fresh context, same as with people. And the retro writes back into the knowledge base, so the next task starts smarter. That last arrow is the whole product.
It doesn’t live in a chat window either. It picks up Jira tickets, works in GitLab branches, merge requests and CI, reads the Sentry stack trace before planning a fix, queries Loki logs around an incident, and when it’s genuinely stuck, asks in Slack and waits instead of guessing. Compared with running Claude by hand, you get memory, process, governance and throughput: a session forgets, a platform compounds.
It’s open source and self-hosted under Apache-2.0, built in public. The knowledge is markdown in your own git repository, so if you delete the platform, the knowledge stays. Implementation has just started. Come break it.
The part nobody puts on a slide
New toys, new problems. Who reviews forty merge requests a day? Throughput moves the bottleneck from writing to reviewing, and review doesn’t scale by hiring one more person. Then there’s AI slop in production: plausible, tested, reviewed, and wrong.
The agent isn’t lying to you; it genuinely can’t tell. Research found that when an agent misjudges its own work, 75.8% of the time it claims success. In the same work, independently verifying the actual state, rather than asking the agent, cut false successes roughly tenfold. That’s why the answer is a pipeline with independent review and deterministic gates, not a bigger prompt.
The adaptation: from engineer to operator
Less typing the solution. More building the context, framing the problem, and knowing when it’s lying.
“Operator” isn’t a demotion. It’s the job that’s left, and it’s the one that actually needs the twenty years.
A friend asked me whether he could build myspeedpuzzling.com with Claude, exactly like I did. My honest answer: yes. But he would need enough time to step in every single mistake I made along the way. Anyone can iterate to the same solution. The real question is how many iterations it takes, and whether you notice you’re wrong. Experience isn’t the code you can write; it’s the wrong turns you no longer take.
Meanwhile, for juniors
This part isn’t funny. Employment of developers aged 22 to 25 is down about 20% from its late-2022 peak, according to the Stanford Digital Economy Lab. The junior share of tech hiring has reportedly fallen from 15% to 7%.
I’m thrilled about what I can do with Claude Code because of twenty years of scar tissue. A junior doesn’t have that yet, and software can replace them easily.
My opinion, again: healthy companies with the budget should invest in juniors. Not as charity, but as compound interest. In fifteen years there will be no seniors if nobody grows any.
Back to the spoiler
Will developers be jobless? Yes. But not you, assuming you’re already uncomfortable. So, how far is your company? What tools are you actually using, what have you built yourselves, and how satisfied are you, honestly? I genuinely want to know.