Three cards, two operators, one structural observation: when agents become the primary authors of code, the humans around them need to update what they optimize for. Ball named the updated selection criteria. Willison named what happens to the people who live inside the updated world.
The selection frame flips when agents write the code
Thorsten Ball's September 26 edition of Joy and Curiosity (issue 101) makes two connected claims about technology selection in an agent-written world.
The first, captured in Language selection criteria shift from developer preference to agent performance when agents write most of the code, is a straightforward implication of who the primary author now is. Developer ergonomics matter while the developer writes the code. Once agents write most of it and developers review and direct, the qualities that predict agent success replace the qualities that made a language enjoyable to type. Ball states the shift in a single sentence, contrasting the old frame ("I enjoy working with it") with the new one ("my agent gets great results with it"). He illustrates the dissolution with a personal data point: three daily production apps running on stacks he could not name if asked.
"Agents are a hundred times bigger than any language, framework, or library choice."
The second card, When agents write code, the decisive technology selection criterion is whether failure modes are legible to the agent, specifies which properties become decisive under the new frame. Agents cannot ask clarifying questions. They diagnose through external signals. Ball names the relevant dimensions: performance characteristics, resource usage, failure modes, observability, debuggability, deployments, rollbacks. These must be legible to the agent through structured output and machine-readable errors, not through conventions that assume an experienced human author reading the code in context. A framework that was pleasant to write in but opaque when it failed becomes a liability in an agent-written codebase. Observability shifts from a quality-of-life concern to a structural requirement.
- Thorsten Ball, Language selection criteria shift from developer preference to agent performance when agents write most of the code. Language selection shifts from developer preference to agent performance metrics.
- Thorsten Ball, When agents write code, the decisive technology selection criterion is whether failure modes are legible to the agent. Observability becomes load-bearing infrastructure when agents are primary authors.
The psychological cost of the shift
Simon Willison's September 27 keynote at WeAreDevelopers World Congress identified a phenomenon spreading through engineering communities that the technology selection discussion does not reach. He named it "Deep Blue," crediting Adam Leventhal and Bryan Cantrill of Oxide Computer for the term.
Software engineers working with capable coding agents report AI-induced ennui called Deep Blue describes the experience of engineers who have already crossed the agent-authorship threshold. The problem is not job loss. Output continues. Throughput increases. The problem is the break in the meaning-loop that previously ran through technical craft. Engineering identity is partly constructed from cracking hard problems. When an agent handles most tractable problems reliably, the engineer stops accumulating that experience. What replaces it, whether directing agents provides the same reward, is an open question that individual engineers are working out for themselves.
The chess parallel holds: after Deep Blue defeated Kasparov in 1997, grandmasters kept playing, but their relationship to the game changed. The full implication of what software engineering is when agents can do most of what engineers do is still forming.
- Simon Willison, Software engineers working with capable coding agents report AI-induced ennui called Deep Blue. Software engineers working with capable agents report a specific professional ennui from the loss of the hard-problem-solving reward.