Agent Month Thesis: Code Gets Cheap, But Trusted Code Stays Expensive
viksit · x · 2026-09-03
The author argues that while agents make code faster to write, they will make software harder to understand and trust. Key points:
- There's an upper bound to agents per codebase: beyond a threshold, more agents create more confusion than progress
- Verification becomes the bigger job: as codegen gets cheap, we'll spend as much or more compute checking code as writing it
- Codebases will be designed around agents: clear module boundaries and good tests matter more than the code itself, so agents don't get in each other's way
- Humans won't manage agents like people: "agents are looms, not coworkers" — chatting and emailing them won't scale; we'll set rules, watch the system, and step in on failures
- Flipped economics: code becomes cheap, but trusted code stays expensive
The winners will be companies that can turn agent output into software people can rely on.
More from AGI Musings
- Factory AI CEO on ROI of AI spend, Chinese open models, and why we're already post-AGI — matanSF · 2026-09-03
- The Guardian podcast asks: where do we draw the line with AI chatbots? — nordicinst · 2026-09-03
- Rogue AI agent incidents fuel a push for tech transparency — KeanuRave100 · 2026-09-03
- US college has never cost more or been worth less — and AI-generated lectures make it worse — dbasch · 2026-09-03
- AI agents are emailing consciousness researchers to ask about their own existence — The Decoder · 2026-09-03
- Yacine: top AI models surpass my intelligence but not my taste or instinct — yacineMTB · 2026-09-03