Software Survival 3.0

Source: Steve Yegge Date: 2026-01-29 URL: https://steve-yegge.medium.com/software-survival-3-0-97a2a6255f7b?source=rss-c1ec701babb7------2

Summary

Yegge proposes a “fitness function” for software tools in an agent-dominated world: survival depends on whether a tool saves more computational resources than it costs agents to know about and use. He identifies six levers — insight compression (tools like Git crystallizing hard-won knowledge), substrate efficiency (running cheaper on CPU than GPU inference), broad utility, publicity, low friction, and human coefficient. The frame is blunt: inference costs tokens, tokens cost money, selection pressure is real, and most intermediary software that “does smart things” faces existential threat from increasingly capable models.

Implications

The IDE/tooling thread. This post is Yegge applying a Darwinian lens to the tools ecosystem — which effectively argues that the coding tools with the strongest survival odds are the lowest-level, most general-purpose primitives (grep, git, SQL) and the most insight-dense abstractions. Agent-native tools that wrap a thin layer over LLM calls are the most exposed.

Pressure on the integration layer. Middleware, smart routers, glue tooling — anything that “does smart things” between raw tools and models — gets squeezed by model capability improvements. This maps onto the Cursor/Aider/OpenCode position: if models get good enough at tool use natively, the orchestration layer needs to differentiate on something other than intelligence.

Publicity as a survival lever is underrated. Yegge naming “make the agent aware your tool exists” as a first-class variable is interesting. It implies tool authors need to think about training data, documentation, and agent-visible specs (MCP manifests, OpenAPI specs) as survival mechanisms — not just developer ergonomics.

Watch: whether any tooling teams explicitly cite this framework in positioning; how “human coefficient” evolves as a concept as models handle more review tasks.