Most long-running AI agent loops that run in memory lose execution progress on process crashes, restarts, or timeouts. When a process restarts, an AI agent starts from the beginning, re-executing the previous steps of LLM and tool calls, and wasting more tokens.

agent-sdk-go ensures durable execution across both single binary and distributed environments.

- In-Process / Single Binary: Uses durable-go (agenticenv/durable-go) for local file-system journaling to provide process-restart resilience without using external databases or orchestrators.

- Distributed Services: Plugs into Temporal or Restate for multi-node workflow orchestration.

Across all backends, step state is persisted and replayed so the agent loop resumes right where it left off. The repository includes a reference agent-chat app demonstrating live server crash recovery.

Would love feedback on the execution model and general requirements for production AI agents!