saved
Towards Self-Driving Codebases
Hraness cites a source capture. The source author remains the source.
gist
Dan Robinson argues engineering orgs are past tokenmaxxing and stuck in the trough: agent armies produced dubious code, not a software tsunami. The next plateau is self-driving codebases—agents owning bugs, prod errors, polish, and growth playbooks—so humans spend time on killer features and architecture. The bottleneck is agent-legible environments and missing primitives (global memory, rot prevention), not model quality. Detail’s path is mine bugs → fix them → use fix traces to prioritize agent-readiness investments.
ideas
- Loops are not the lasting job. Viable agent loops are hard today because the toolchain is unfinished; once primitives exist, loops get cheap and trusted, and engineering time shifts to having good ideas.
- Self-driving owns the low-level surface. Bug detection and fix, prod debugging, agent prompt optimization, design-system coherence, QoL polish, and growth experiments should run on GPUs while humans keep feature and architecture judgment.
- Missing primitives are table stakes. Agent-legible envs, cross-tool global memory of corrections, and codebase-rot prevention sit beside evolved APM, CI, and agent-operable A/B testing.
- Environment beats harness right now. Agents fail where they cannot see or reproduce; blind spots cut trust and force human review. Hillclimb agent readiness by mining and fixing real bugs, then closing the env gaps those traces expose.
- Factories disappoint without that science. Codebase-specific agent-ready schlep is still an art; turning it into measurable investment is how software factories start humming.
quotes
“At this point the limiting factor for dev agents is the environment they operate in, not the models and harnesses themselves.”
“Rather, the most valuable engineering work is going to be having good ideas.”
“Where agents have blind spots, they make mistakes, which limits trust”
“That’s one reason so many software factories are so disappointing.”