Skip to content
MAESTRO

Production Mode

Production mode turns a one-off generation into a durable, measurable, improvable agent. It’s the path you use when you want to keep a chain, test it, evolve it, and ship the best version.

What happens on a Production prompt

Section titled “What happens on a Production prompt”

When you type a task in Production mode, MAESTRO runs this sequence automatically:

  1. Generate — MAGE produces a reproducible chain (no ReAct loop, no answer-synthesis — Production chains must run the same way every time).
  2. Stash — the chain lands in the session artifact store (the header pill and /artifacts see it).
  3. Save — the chain is saved to Memory under a stable chain_id with a display name. (Duplicate of an existing chain → no re-save.)
  4. Baseline — MAESTRO runs one baseline execution and persists it as the first dataset entry for that chain.
  5. Evolve — if a Platform is wired and the baseline succeeded, MAESTRO kicks off an evolution run against the baseline.

After this, the chain lives in your Library under its chain_id.

Requirements & fallback

Section titled “Requirements & fallback”
  • Memory is required: CARE_MEMORY__BASE_URL (+ CARE_MEMORY__API_KEY if auth is enforced). Without it, selecting Production auto-falls back to Ad-Hoc with a warning.
  • Platform is optional: evolution only runs when CARE_PLATFORM__BASE_URL is set; otherwise the save + baseline still happen, evolution is skipped.

Channels & versions

Section titled “Channels & versions”

Saved chains are versioned. Edits (via /revise) create new versions; the latest channel always points at the newest. Promote a chosen version (or an evolution winner) into the stable channel with /promote. CLI reads honour --channel (default latest) — e.g. care run <id> --channel stable.

Production commands

Section titled “Production commands”

These appear in Production mode (see the full list under slash commands → production):

Datasets — measure quality

Section titled “Datasets — measure quality”
/dataset add <chain_id> "<task>" --expected "<out>" [--rubric "<prompt>"]
/dataset list <chain_id>
/dataset run <chain_id>          # replay every entry + score it
/dataset export <chain_id> <path>   # write entries as JSONL

The baseline run seeds entry #1; add more cases, then /dataset run to score the chain against them. The CLI twin builds + scores datasets headlessly.

Evolution — improve automatically

Section titled “Evolution — improve automatically”

You can launch an evolution run three ways:

  1. Automatically — in Production, a successful baseline kicks one off (step 5 above).
  2. From the TUI — open a saved chain in the Library and press v / E to open the Evolution Launch picker. It shows a budget preview before you commit — the iteration count, rubric, and objectives, with the estimated token/USD cost for the run — so you can size a run before spending on it.
  3. Headless — care evolve <chain_id> --iterations 8 --wait --accept.

Once a run is live, follow it on the Evolution dashboard — the list of active and recent runs (Enter opens one, c compares two). Each run view streams the Fitness chart, the Pareto front, Programs, and Versions, with a running cost meter in the header so you always see token/USD spend against the budget you set at launch.

Watch and steer a run from chat:

/evolution <run_id>              # render the run's state inline
/evolution watch <run_id>        # stream events live
/evolution accept <run_id> <individual_id>   # promote the winner

The GA optimises against the chain’s dataset (the fitness signal), so seed a few dataset entries first. Accepting a winner — via /evolution accept, the Evolution dashboard’s Accept winner button, or care evolve --accept — promotes the best individual into the stable channel.

Deploy — serve as an HTTP agent

Section titled “Deploy — serve as an HTTP agent”

Ship a saved chain to the agent hub so it runs as an HTTP agent with its own Swagger UI:

/deploy <ref> [--channel <ch>] [--name <agent>]   # default channel: stable

Needs the deploy extra (pip install "maestro-care[deploy]"). Full guide: Deploying agents.

Lifecycle

Section titled “Lifecycle”
/revise [<id>] <change>          # edit the chain in natural language → new version
/promote <id> <version>          # promote a version / winner to the stable channel
/upload <chain_id>               # POST the chain to CARE_UPLOAD__URL
/forget <chain_id> [--force]     # soft-delete the chain + its dataset

See also

Section titled “See also”