Agents ace isolated tasks but drop to ~38% at maintaining codebases over time
FullStackFrenzy · reddit · 2026-11-05
A Reddit post engages with Zhenfeng Cao's paper "The End of Software Engineering" (arXiv:2606.05608), which argues code becomes disposable and humans become "intent architects". The poster counters with the paper's own evidence:
- On EvoClaw, agents score 80%+ on isolated one-off tasks
- But they fall to at most 38% when they must keep a codebase working across changes, because early mistakes carry forward
- The real worry isn't broken builds but slow correctness drift: a billing feature where tests pass, then weeks later refunds quietly fail for one provider that retries webhooks twice
The takeaway: writing code got cheap, keeping a system correct over time didn't. The author asks what's catching slow drift for practitioners — evals, a separate verifier model, or plain prod monitoring.
More from coding & agent
- Keynote built entirely with OpenClaw agent via voice notes on a train ride — vincent_koc · 2026-11-05
- DeepSeek's dsh-superagent: a coding agent whose rulebook is 'stop narrating your work' — shoumikgoswami · 2026-11-05
- OpenRouter vs OpenAI Zero Data Retention for production: a developer's five open questions — kshitizsriv · 2026-11-05
- MetaMask launches Agent Wallet with onchain guardrails for AI agent spending — LexSokolin · 2026-11-05
- Dev questions Decision RAG: why not just use Jev embeddings for vector search — solyarisoftware · 2026-11-05
- Soniox launches STT v6 with agent text input, first STT aware of both conversation sides — solyarisoftware · 2026-11-05