# 2026: The Year the IDE Died

Steve Yegge, Sourcegraph/Amp & Gene Kim, IT Revolution | AI Engineer CODE 2025 | 24:59

Source: https://www.youtube.com/watch?v=7Dtu2bilcFs
Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks.
Page: https://aietalks.com/talks/2026-the-year-the-ide-died
Published: 2025-12-06
Tags: agents, code-generation, engineering-culture, multi-agent, team-adoption

## TL;DR
- Coding agents are too difficult and risky for most developers, so future tools will use interfaces, task decomposition, and multiple specialized agents.
- AI-assisted coding is changing who can build and ship software, allowing leaders, support teams, and domain experts to create production features without typing code by hand.
- Adopting these tools requires sustained practice because trust in AI coding systems grows with time and experience.

## Summary
Steve Yegge argues that current coding agents, including Claude Code, are powerful but too difficult and risky for most developers. He expects coding tools to move from command-line agents toward interfaces that coordinate many specialized agents, much like a CNC machine controls precise work. Gene Kim then connects this shift to changes in technology organizations. He defines vibe coding as an iterative conversation in which AI writes the code, and describes how it makes ambitious or tedious work easier to attempt. Examples from Booking.com, Traveloka, Fidelity, Cisco Security, and a workshop for technology leaders show people outside traditional development teams shipping software. Kim also discusses research suggesting that trust in AI tools increases with continued use. The talk is optimistic about what smaller teams and non-developers can build, while it admits that organizations still lack good ways to coordinate large volumes of AI-generated code.

## Key ideas
### Current coding agents are powerful, but their cognitive overhead blocks adoption
[00:36](https://www.youtube.com/watch?v=7Dtu2bilcFs&t=36s)
Yegge says developers are not adopting Claude Code and similar tools because they are too hard to use. He compares Claude Code to an electric drill or saw: an untrained person can cause serious damage, while a skilled craftsperson can do precise work. Software makes the problem larger because both the codebase and the ambition can be effectively limitless. He also says agents can "lie, cheat, and steal," which makes many developers uncomfortable. His argument is that better models alone will not solve the adoption problem. The tools need to reduce the amount of judgment and operational effort required from the person using them.

### Coding tools will move from manual agents toward controlled machine systems
[01:58](https://www.youtube.com/watch?v=7Dtu2bilcFs&t=118s)
Yegge predicts that the next generation of coding tools will resemble CNC machines. A CNC machine takes coordinates and moves a tool precisely, while future coding systems will let engineers specify what should happen without directly inspecting every line of code. He rejects the claim that progress stops if models plateau, saying the industry has already discovered the equivalent of steam and electricity and now faces an engineering problem. He predicts that within a year or a year and a half, code will be written by large automated systems overseen by engineers who no longer look at the code directly.

### The new coding interface will coordinate many agents instead of enlarging one agent
[04:17](https://www.youtube.com/watch?v=7Dtu2bilcFs&t=257s)
Yegge says the next tool should have a user interface rather than look like an IDE or command-line program. He praises Replit as one company moving in that direction. He also criticizes current agents for behaving like one enormous ant that sends every request to an expensive model. His diver metaphor explains the context problem: a single agent sent into a codebase will eventually run out of its context window, just as a diver runs out of oxygen. A product-manager agent should first define the work, followed by coding, review, testing, and merge agents. He expects systems to use task decomposition, successive refinement, components, and black boxes.

### AI-assisted coding is creating a large productivity gap inside engineering organizations
[03:00](https://www.youtube.com/watch?v=7Dtu2bilcFs&t=180s)
Yegge recounts Andrew Glover describing a split at OpenAI between engineers who use Codex and those who do not. He says the productivity difference is so large that it creates a problem during performance reviews, because people with the same title may produce vastly different amounts of work. He compares the situation with the Swiss mechanical watch industry after quartz technology appeared. In his account, senior and staff engineers are among those refusing the new tools. He presents this resistance as an organizational problem rather than an individual preference, because managers may soon have to compare engineers working with very different levels of automation.

### Vibe coding lets people outside development teams build and ship features
[10:22](https://www.youtube.com/watch?v=7Dtu2bilcFs&t=622s)
Kim defines vibe coding first as anything where a person does not type code by hand, then adopts Dario Amodei's definition: "the iterative conversation that results in AI writing your code." He says the practice makes impossible projects possible and makes tedious tasks cheap. A Cloud Code team example describes fixing customer issues immediately instead of putting them through a Jira backlog and lengthy grooming process. Kim also says people can work more autonomously because they no longer need to wait for another team, and markdown files plus language models can improve coordination between specialties. He describes vibe coding as faster, more ambitious, more autonomous, fun, and useful for creating more options.

### Small teams are replacing larger groups for some software projects
[17:36](https://www.youtube.com/watch?v=7Dtu2bilcFs&t=1056s)
Kim gives several examples of smaller groups using AI to build software. At Booking.com, an experiment across 3,000 developers produced a double-digit productivity increase, with faster merges and shorter peer-review times. At Traveloka, a legacy application was replaced in six weeks by a very small team. Kim says the earlier pattern might have required six developers, a UX person, and a product owner, while the newer pattern might use a developer and a domain expert. At Fidelity, Topobroto Pal built in five days an application that his team had estimated would take about five months and require a front-end hire. A junior engineer then maintained it, while usage grew tenfold.

### Trust in AI coding tools grows through sustained practice
[21:11](https://www.youtube.com/watch?v=7Dtu2bilcFs&t=1271s)
Kim discusses an unpublished finding from the DORA research about how long people have used AI tools and how much they trust them. He defines trust as the ability to predict how another party will act and react. With more trust, people can make larger requests, use fewer words, and need less feedback. He says the research showed that people who use AI tools longer tend to trust them more. That makes quick judgments such as "I tried it and it's terrible at coding" less meaningful when the trial lasted only an hour or two. Kim presents effective AI use as a teachable skill that improves through frequency, intensity, and practice.

### Organizations still lack a way to coordinate very high volumes of generated code
[23:23](https://www.youtube.com/watch?v=7Dtu2bilcFs&t=1403s)
Kim closes with examples of leaders and non-developers shipping software, including a workshop where every participant built something in three hours. One participant built an iOS app and submitted it to the App Store review queue. Kim also shares organizational reactions to AI-generated code. A technology leader said colleagues reacted badly after hearing that AI had written 60,000 lines of code that the leader had not inspected. In another case, senior engineers accepted an AI-generated fix for a decade-old problem after previously rejecting AI-generated work as "AI slop." Kim says code velocity has become so high that he and Yegge have concluded they can only have one engineer per repository because of merge conflicts.

## Notable quotes
- "Claude code ain't it. Completions wasn't it. I love Claude code. I use it 14 hours a day. But it ain't it." (00:55)
- "All code within a year, a year and a half, will be written by giant grinding machines overseen by engineers who no longer actually look at the code directly anymore." (02:39)
- "Vibe coding is really the iterative conversation that results in AI writing your code." (12:27)
- "We are probably going to be the last generation of developers to write code by hand. So let's have fun doing it." (13:01)
- "Our code velocity is so high, we've both concluded that we can only have one engineer per repo, because of merge conflicts." (23:59)

## Tools & references mentioned
- Claude Code
- Codex
- Replit
- OpenAI
- Sourcegraph
- Amp
- Anthropic
- Dario Amodei
- Andrew Glover
- Brendan Hopper
- Gene Kim
- Steve Yegge
- Jordan Hubbard
- Nvidia
- Tripwire
- Jeff Bezos
- Amazon
- Rick Dalzell
- The Wall Street Journal
- Adrian Cockcroft
- Netflix
- AWS
- Zapier
- The Phoenix Project
- State of DevOps Research
- DORA
- Google Cloud
- Nicole Forsgren
- Jez Humble
- Booking.com
- Bruno Pasqualis
- Traveloka
- Shri Balakrishnan
- Fidelity
- Topobroto Pal
- Neo4j
- Log4j
- Cisco Security
- John Routhier
- Eric Meijer
- Visual Basic
- C#
- LINQ
- Haskell
- Hack
- Meta
- Southwest Airlines
- Vibe Coding

## Who should watch
- You manage engineers and need to understand why some people adopt coding agents while senior developers resist them.
- You are a developer deciding whether to move from an IDE and manual coding toward agent-based workflows.
- You lead a support, product, design, or operations team and want examples of non-developers using AI to ship software.

## Related talks

- [Developer Experience in the Age of AI Coding Agents](https://aietalks.com/talks/developer-experience-in-the-age-of-ai-coding-agents) (Max Kanat-Alexander, Capital One, 18:20)
- [Everything We Knew About Software Has Changed](https://aietalks.com/talks/everything-we-knew-about-software-has-changed) (Theo Browne, @t3dotgg, 16:02)
- [Building an Autonomous Engineering Org](https://aietalks.com/talks/building-an-autonomous-engineering-org) (Angie Jones, Agentic AI Foundation, 17:36)
- [Software Fundamentals Matter More Than Ever](https://aietalks.com/talks/software-fundamentals-matter-more-than-ever) (Matt Pocock, 18:26)
- [AI Leadership](https://aietalks.com/talks/ai-leadership) (Alex Lieberman, Morning Brew and 10X & Kath Korevec, Google Labs & Katelyn Lesse, Anthropic & Michele Catasta, Replit & Lisa Orr, Zapier & Steve Yegge, Sourcegraph and AMP & Gene Kim, IT Revolution & Bill Chen & Brian Fioca, OpenAI & Martin Harrysson & Natasha Maniar, McKinsey & Yegor Denisov-Blanch, Stanford & Itamar Friedman, Qodo & Olive Song, MiniMax & Asaf Bord, Northwestern Mutual & Lei Zhang, Bloomberg & Samir Mody, The Browser Company & Max Kanat-Alexander, Capital One & Arman Hezarkhani, 10X & Justin Reock, DX & Dan Shipper, Every & Mel Lutzky, Graphite, 8:16:05)