Anthropic's head of platform engineering: long-running agents create new engineering problems
every · x · 2026-11-05
Katelyn Lesse, head of platform engineering at Anthropic, argues software engineers will matter more, not less, in the automation era. Thousands of agents working for weeks on hard problems create a new class of engineering challenges:
- Infrastructure: agents are isolated, single-instance workloads that idle for long stretches then must start quickly
- Permissions: who decides what agents may do, and how do humans meaningfully approve actions in multi-agent systems running for weeks?
- Security & isolation: vulnerabilities surface faster than they can be patched; thousands of persistent agents reading/writing untrusted content makes them far more dangerous
- Monitoring: we need systems that can identify dangerous behavior
Solving these is existential for AI to deliver its potential — and fun for engineers.
More from AGI Musings
- Stanford's James Zou on AI agents entering the wet lab, in Nature — james_y_zou · 2026-11-05
- Pew: 60% of Americans uncomfortable with doctors using AI, only 3% accept AI-only diagnosis — rohanpaul_ai · 2026-11-05
- Quanta: Injecting 'Thoughts' Into LLMs to Test Whether AI Can Introspect — mtizard · 2026-11-05
- Most new users can't figure out general-purpose chat agents — use cases are too broad — signulll · 2026-11-05
- AI writes, AI reviews: peer review broken by identical LLM reports — Dr_Alex_Crimi · 2026-11-05
- LLMs now beat human superforecasters at predicting the future — jzlegion · 2026-11-05