Durable AI agents
- Problemstep 1
- Agent runs take minutes to hours, make dozens of LLM and tool calls, and lose all progress when a process dies.
- Patternstep 2
- Run the agent loop as a workflow, make each LLM call and tool execution an activity, and use signals or updates for human approval.
- In the wildstep 3
- Emergent runs 10 to 30 minute agent builds with sub-agents as child workflows.


