2026: The Year of the Uber Engineer

One of my favorite movies of all time is V for Vendetta. There’s a particular scene that sticks in my memory: it’s towards the end of the movie, when Inspector Finch is talking to his assistant Dominic Stone. Here’s the clip (it's worth the 1-minute watch).

This is how I feel about AI in 2026, specifically AI-assisted software development. Let’s start with a timeline of the last 18 months. These are just the major milestones that I’ve personally engaged with in a meaningful way; trying to provide a timeline of all AI launches over the last 18 months would make this post insanely long.

Some quick points first:

1. I reference a lot of Steve Yegge’s posts and work. In my view, he’s the current thought leader in AI-assisted development and literally wrote the book on Vibe Coding. He’s also pushing the envelope with the tools he’s building, as you will see later in the post.
2. While this post was AI-assisted, of course, the vast majority of it was written by me.

Let’s begin!

Timeline

June 2024

Steve Yegge writes the Death of the Junior Developer, where he introduces the concept of Chat-Oriented Programming (CHOP). He argues that software engineers who refuse to adopt CHOP as their primary workflow are becoming obsolete. He calls out (quite early) that programming has moved from manual, line-by-line coding to a "supervisory" role where developers guide and review AI-generated code.

December 2024

Steve follows up on his earlier post with Death of the Stubborn Developer. Here, he clarifies that it’s not junior developers who will become obsolete, but rather people who refuse to adopt the new way of working. He notices that many senior developers are stubbornly refusing to leverage AI to help them write code.

February 2025

Claude 3.7 Sonnet and Claude Code are released.  This is a monumental moment in the world of AI-assisted programming. We go from largely copy-paste between things like ChatGPT and Claude to a true AI assistant that can operate autonomously as an agent. The model also feels smart. It only gets 62% on SWE Bench Verified, gets stuck a lot, and does stupid things, but it’s still a massive shift in how software is built. This is the first real “I feel a disturbance in the force” moment for many people.

March 2025

Steve writes Revenge of the Junior Developer. The "revenge" is that junior developers are adopting Agentic Workflows (AI that acts on its own) much faster than senior developers, who are often stuck in "stubborn" traditional habits.

Yegge outlines 6 Waves of AI development:

1. Manual Coding: (The "Barbarian" era).
2. Completions: (Copilot/IntelliSense).
3. Chat (CHOP): (Iterative prompting).
4. Agents: (AI that uses tools, runs terminal commands, and fixes its own errors).
5. Clusters: (Managing 3–5 parallel agents).
6. Fleets: (Managing massive swarms of agents).

His core argument is that while senior devs are still debating whether "code completion" is helpful, juniors are jumping straight to Wave 4 (Agents). Because they don't have decades of manual habits to unlearn, they are becoming "Agent Managers" who can outperform senior devs still writing code by hand.

May 2025

Claude 4 is released, marking another step change in model capability. It gets ~73% on SWE Bench Verified, and it becomes clear that these models will continue to improve significantly every ~3 months.

June 2025

Steve’s latest is The Brute Squad, where he argues that we are moving from single-agent workflows to orchestrating many agents simultaneously, allowing a single engineer to perform like a team of old-school devs.

September 2025

I discover Conductor via a coworker, which feels like the first pleasant UX for running multiple AI agents in parallel. Still, there is a lot of pain around merge conflicts, and while this is a glimpse at the future, it’s evident that tooling needs to improve.

Later in the month, Anthropic releases Claude 4.5 Sonnet. This one feels more incremental, going to 77% on SWE Bench. Still, it emphasizes that progress will not stop.

October 2025

Steve launches Beads. It gives AI agents memory and quickly becomes a critical piece of scaffolding for engineers serious about using AI to build production software and run multiple agents.

November 2025

At KOHO, we start using Figma Code Connect in earnest. We are behind the times on this one, to be honest, since the tool was released in 2024. This is the first time it feels useful for our specific use cases. Developers say that a project that used to take 1-2 weeks (even with AI assistance) was done in 2 days with Code Connect.

This same month, Claude 4.5 Opus launches with reduced pricing. We switch everyone to this model the day it comes out, and with an 81% on SWE Bench, it takes the crown as the most capable model for AI-assisted development.

December 2025

Adam Wolff, who leads the Claude Code team at Anthropic, writes The End. His key thesis is that coding is "done," but Engineering is just beginning. The central argument is that the mechanical act of "coding" (translating human intent into computer syntax) is effectively a solved problem thanks to models like Claude 4.5.

January 1st, 2026

Steve launches Gas Town. It’s rough around the edges, but it’s incredibly ambitious in its attempt to solve all of the challenges with coordinating multiple AI agents in one fell swoop.

So What Happens Now?

I don’t know, but I can guess. We know the models will continue to improve and get cheaper. Claude 5 is almost certainly coming out in Q1, with another performance bump. As Yegge calls out, the foundational LLMs will start to have Beads, Gas Town, and other tools in their training data, making these tools even more effective. Gemini is catching up to Claude quickly, and while I’m not going to place a bet on the best coding model, one of them will be much better in 2026 than any model in 2025.

Tools like Beads and Gas Town will get dramatically better (or new tools will rise to displace them). Merge conflicts will no longer be a concern - the tooling will handle it for you. You won’t have to re-prompt AI that gets stuck; a different AI agent will detect it and unstick the stuck AI for you. People complain that their lives are turning into code review, but that, too, will be automated away. Why have a human read code when you can have several agents review it, create a robust test harness, compare notes, and create a far superior review to anything a human could do?

Teams will shrink dramatically. We’ve already reduced the atomic unit of a pod at KOHO from 6 devs to 4, and it’s plausible that by the end of 2026, a “pod” will be one dev and a product manager. Related, the roles of product manager and UX designer will merge. So much of a PM's work can be done by AI that you won’t want two people in the flow (every handoff adds latency to the process of shipping software). UI designers will exist, largely building a pattern library and setting the design language for a company. With this foundation, AI will generate all feature UIs and all code.

Spec-Driven Development (SDD) will take over. The critical skills for building working software will become:

1. Deeply understanding the customer problem to be solved
2. Having the ability to translate this problem into requirements
3. Being able to articulate these requirements to an LLM and iterate with the LLM to get to the outcome you want.
4. Coordinating many agents simultaneously to get to this outcome

The differentiating skillsets for an engineer will be empathy, interpersonal skills (to work with your product counterpart), high cognitive throughput (to maintain state awareness across multiple parallel workstreams), and taste. Knowing when something feels right, architecturally, UX wise, and visually is key.

It will take time to evolve towards this state. Early stage startups are already there, and big companies will take years. As always, failing to evolve and adapt to the new reality will mean displacement by a better, cheaper alternative.

People and companies who DO adapt will be able to ship software so quickly that it will blow people’s minds. Full-featured, complex apps in a few short weeks. We saw glimpses of this in 2025 - Steve Yegge built Gas Town in 17 days. OpenAI built Sora for Android in 18 days, with a public launch in 28 days.

Conclusion

There’s still a surprising amount of pushback against AI, and rage against its coding abilities. We see this even from some of the best engineers in the world, like Rob Pike. There is a valid fear that we are losing the 'craft' of programming, but I would argue the craft is simply evolving. Rob Pike is mourning the loss of the solo woodworker, while we are busy inventing the assembly line. The 'soul' of the machine hasn't disappeared; it has just moved up the stack to where the strategy and architecture live. 

Ultimately, the future of software engineering isn’t about fighting the tide of automation; it’s about learning to ride it. We are witnessing the birth of the Uber Engineer.

The Uber Engineer isn't the person who can write a red-black tree in C++ from memory. They are the "Agent Manager" who can maintain state awareness across many parallel workstreams, ensuring their fleet of AI agents isn't just writing code, but solving business problems. They focus on the high-leverage human skills that no LLM has mastered: empathy for the user, impeccable architectural taste, and the judgment to know when to embrace the LLM's response, and when to throw it away and start over. 

The next few years promise to be transformative. The mechanical act of coding may be done, but there is a tremendous amount of software to build and problems to solve. Those who lean in and embrace the role of the Uber Engineer won't just keep their jobs—they will define what it means to build the future.

If you want to lean in and be at the cutting edge of software development in 2026, KOHO is, of course, hiring.