Living with a Long-running Personal Agent: Workflows, Boundaries, and Failure Lessons
This presentation works best with JavaScript, but this no-JS outline preserves the core flow, sources, and public-safety boundary. Build with npm run build.
- Opening: Anthropic workflow-vs-agent distinction, plus talk-local chatbot → deterministic workflow → agent sequence.
- Two-axis Personal Agent map: action scope × continuity, concrete dated configurations, no Atom highlight.
- Model ≠ agent: harness, state, tools, policy, workers, verification.
- Public-safe Atom case: why I need continuity and coordination.
- Boundaries/security and three failure lessons: memory, wrong GUI state, false completion.
- Takeaways and source cards.