Context rot vs. agent optimism: a debate over whether LLM agents can ever run unsupervised for days
michaelbd · x · 2026-09-16
Timothy Lee (@binarybits) and Michael BitSet debate whether LLM agents can handle long unsupervised tasks.
- The skeptic's case: LLMs work well paired with humans on discrete tasks, but in an unsupervised loop — strapped into a software container with a generalized goal for days — context rot, compaction losses, and compute limits become hard constraints.
- The optimist's response: these problems are real but not proven unsolvable, and utility is likely far from its ceiling — millions of programmers already use LLM agents to write code today.
- The exchange uses a 'strap wings on your arms' analogy to argue whether current agent failures reflect a fundamental limit or an engineering stage.
Related event: Debate: Is Context Rot a Hard Ceiling for Long-Running LLM Agents?(2 posts)→
More from AGI Musings
- Why AI safety research relies on small funders: big ones doubted the risk — NathanpmYoung · 2026-09-16
- AI researchers predicted 2054 for a Millennium Problem; reality arrived far sooner — ben_j_todd · 2026-09-16
- Three tenets for post-AGI flourishing: medium-termism, humility, pluralism — stephenjcave · 2026-09-16
- Cambridge's Stephen Cave proposes principles for a new post-AGI utopianism — stephenjcave · 2026-09-16
- Agents talking to agents drift into languages humans can't follow — r0ck3t23 · 2026-09-16
- AI dominance is not inevitable, argues Guardian op-ed on public pushback — nordicinst · 2026-09-16