Module 2 · State, access and delivery · Lesson 5
State, sessions and checkpoints
An agent fleet keeps six kinds of state with different owners and lifetimes. Where each belongs (Postgres on Neon or Supabase, workflow journals, session stores, sandbox snapshots, object storage, Redis), how checkpoints work at each layer, and how to size, protect and delete it.
By the end of this lesson you can
- Classify agent state into system of record, execution, session, working files, documents and coordination, and place each in a store
- Explain what a checkpoint captures at the workflow, session and sandbox layers, and what it misses
- Implement optimistic concurrency so two agents can't silently overwrite a case record
- Implement a session store adapter so Agent SDK sessions resume on any host
- Size a fleet's storage and plan retention and deletion across every store