From IDEs to AI Agents with Steve Yegge

Timestamps

The 8 Levels of AI Adoption

Steve's framework for where engineers sit with AI tools. His claim: 70% are stuck at levels 1-2.

flowchart TD
    L1["Level 1: No AI"] --> L2["Level 2: Yes/No — basic Q&A"]
    L2 --> L3["Level 3: YOLO — trust is rising"]
    L3 --> L4["Level 4: Squeeze — watching conversation, not diffs"]
    L4 --> L5["Level 5: Agent-only — IDE is secondary"]
    L5 --> L6["Level 6: Multi-agent — bored, fire up another"]
    L6 --> L7["Level 7: The Mess — coordination problems"]
    L7 --> L8["Level 8: Orchestration — agents running agents"]

    L2 -.- note1["70% of engineers stuck here"]

The jump from level 6 to 7 is where things break — you accidentally text the wrong agent, changes conflict, coordination falls apart. That's exactly the problem Gas Town (Steve's open-source orchestrator) tries to solve. The pattern: completions → chat (completions in a loop) → agents (chat in a loop) → orchestrators (agents in a loop).

Key Arguments

The Abstraction Ladder Never Stops

Steve's 40-year view: what engineers "need to know" keeps shifting upward. Assembly → compilers → high-level languages → distributed systems → AI. He watched graphics go from pixel-plotting to game-world physics in a decade. The same compression is happening to software engineering now. His own ego was wrapped up in compiler knowledge — and he had to let it go.

Big Companies Are Quietly Dying

Google hasn't produced organic innovation since ~2008 (only acquisitions). When a company has more people than work, politics and empire-building kill creativity. Steve's prediction: AI-enabled teams of 2-20 people will rival or exceed the output of large enterprises. "We're all looking at the big dead companies. We just don't know they're dead yet."

The Vampiric Effect Is Real

AI makes you extraordinarily productive but drains System 2 thinking at an unsustainable rate. You might get 3 genuinely productive hours at max vibe coding speed — and still produce 100x output. The question becomes: who captures the surplus? If companies demand 8-hour days at AI speed, people break. Steve wrote a whole article about this — the-ai-vampire expands on the economics and the pushback strategies.

Monoliths Are Incompatible with AI Development

The ceiling for what agents can productively build is between 500K and 5M lines of code — and rising. Enterprise monoliths with hundreds of millions of lines will never fit in a context window. Companies must break them up or consider rewriting from scratch, which is now becoming faster than incremental migration.

"Heresies" Are the New Tech Debt

In vibe-coded codebases where humans don't read the code, agents can propagate incorrect architectural decisions. Steve calls them "heresies" — they spread, recur even after removal (a single stale doc reference can resurrect them), and must be explicitly guarded against in system prompts. This connects directly to the bitter lesson: don't try to be smarter than the AI, but you still need to constrain it.

Notable Quotes

"I along with everyone else went 'no, it's not' — just flat-out rejection, absolutely not happening — until I used it and then I was like, 'Oh, I get it. We're all doomed.'" — Steve Yegge on first using Claude Code

"You might only get three productive hours out of a person at max vibe coding speed, and yet they're still 100 times as productive as they would have been without AI. So, do you let them work for 3 hours a day? And the answer is, yeah, you better, or your company's going to break." — Steve Yegge

"We're all looking at the big dead companies. We just don't know they're dead yet." — Steve Yegge on big tech innovation

"I was checking off things that no longer mattered that I had really cared about — my ability to memorize, my ability to write, my ability to compute... the world goes monochrome, all color disappears." — Steve Yegge on the grief of losing craft identity

Predictions

• ~50% of engineers at big companies will be cut — companies set a dial, shed half their engineers to fund AI tokens for the remaining half
• Programming by talking to a face by end of 2025 — most people will program by speaking to an AI avatar rather than typing
• Non-technical family members as top contributors by summer 2027 — Steve predicts his wife will be the top contributor to their game project
• At least two more capability cycles remain — models will become at least 16x smarter than current state
• Gas Town itself will be obsolete in ~6 months — the orchestrator shape that works now won't be the shape that works in mid-2026

The Minimax Context Argument

Two opposing schools Steve identifies: Maximizers fill the context window with rich context for wise responses. Minimizers want shortest possible windows to avoid quadratic cost increases and cognition drop-off. Gas Town uses both — "crew" roles for max context (design discussions) and "polecats" for min context (well-specified subtasks).

Connections

• the-ai-vampire — Steve's own article expanding on the vampiric burnout effect he discusses here, with the economics of value capture between employers and employees
• 2026-the-year-the-ide-died — Steve's talk at AI Engineer 2025 making the same case about IDE death and agent swarms, now elaborated with the 8 levels framework
• boris-cherny-creator-of-claude-code-pragmatic-engineer — Same podcast, Boris built the tool that Steve says changed his mind about AI coding ("until I used it and then I was like, oh, we're all doomed")
• from-tasks-to-swarms-agent-teams-in-claude-code — Alexander's practical experience with multi-agent coordination maps directly to Steve's levels 6-8 and the "mess" problem orchestrators solve