Magistry is a roster of 14 agents, not a monolith. Each agent owns one surface and writes to the same decision_log; the three deep agents — Catalog, Campaign, and CS — additionally run the full planner-judge-executor loop. Marketing names below, in-app names in parentheses where they differ.
The three deep agents
- Catalog Specialist (in-app: Catalog Agent) — owns SKU lifecycle, with seven specialist skills. Classifies products into states, proposes transitions gated by ALLOWED_TRANSITIONS, and runs the margin-aware moves that keep the catalog earning its shelf space.
- Campaign Specialist (in-app: Campaign Manager) — owns paid acquisition, with four specialist skills. Budgets, bids, negatives, and dayparting, executed on Google Ads; Meta, TikTok, Pinterest, and Bing are reached through the ad connectors and feed syndication. Bids on profit rather than revenue where costs allow it.
- CS Specialist (in-app: CS Agent) — owns the inbox, with 8 specialist skills. Reads, classifies, and drafts replies to inbound support, with refund and replacement actions inside guardrails. Its Reply Judge scores every draft on policy grounding, factual grounding, and brand voice, and holds anything below a per-dimension floor for a human.
The supporting cast
- Growth Scout — researches product opportunities and demand signals across sourcing lanes, so the next product decision starts from evidence.
- Copy Crafter — rewrites product copy and runs the 14-day A/B tests that decide whether a rewrite ships or reverts.
- Promotions Strategist — plans and runs discounts inside the margin floor, so no promotion can price a product into a loss.
- Market Expert — a signals agent that gathers credibility-scored market intelligence — competitor moves, category trends, pricing context — and feeds it to the Campaign Manager.
- Storefront Conductor — orchestrates theme and storefront changes, keeping presentation moves as reversible as catalog ones.
- Merchant Guardian — watches Merchant Center for feed disapprovals and auto-fixes the mechanical classes: policy, attributes, formatting.
- Supplier Sentinel — keeps product costs current and flags supplier stock risk before it becomes a stockout or a margin surprise.
- Daily Reporter — writes the daily operating summary: what happened, what changed, what needs a human.
- Customer Pulse — monitors reviews and mentions, surfacing sentiment shifts before they show up in the numbers.
- Storefront Auditor — continuously checks the storefront for broken flows, dead links, and conversion blockers.
- Self-Calibrator — learns winner and loser thresholds from your own sales distribution instead of industry rules of thumb, and re-anchors them as the store changes.
One supervision model for all of them
No agent on this roster is a special case. The three deep agents run the full planner-judge-executor loop; every agent, deep or supporting, is gated by the same policy checks. Everything every one of them does lands in decision_log with its evidence, and reverts are derived from the logged action. And the kill switch covers all of them: one flag, and every job that writes or spends refuses to start. The roster grows; the rules don't bend.