Task-oriented recipes for building agent systems on Strata. Each one is a short sequence of real commands with the exact output you should see. Follow the steps in order and you land on the same results.

Every recipe writes to a durable database directory (for example ./ab) that each command reopens. The --cache flag is single-process only — separate invocations do not share an in-memory database — so multi-step recipes keep the database on disk.

• A/B Testing with Branches — fork one branch per variant, run each in isolation, compare the results.
• Agent State Management — hold config in KV, working memory in a JSON document, and an action log in events; inspect earlier state with versioned reads.
• Deterministic Replay — record external inputs in the event log and reconstruct any past state with versioned reads and fork-at-version.
• Multi-Agent Coordination — give each agent an isolated branch, share an append-only event journal, and separate runs with spaces.
• RAG with Vectors — store embeddings in a vector collection and source text in KV, then query and fetch the matching rows.