Appearance

Agent — user guide ​

AGENT-P9-04 (#1991). How Kasia/Magda/Marcin use the PIM agent. Operator/BYOK procedures are in docs/operations/agent-runbook.md.

What the agent is ​

A conversational assistant for the catalog. You describe what you want in plain language; the agent grounds it in real numbers, proposes a change, and nothing is written until you approve it. It works within your permissions — it can only do what your role lets you do.

Two ways in (same backend) ​

  • Chat panel — the Sparkles icon in the top bar. Best for longer, multi-step work ("prepare the DE launch for category X") and back-and-forth.
  • Cmd+K — from any list. Best for a quick command carrying what you're looking at (the active filter, your selection). Free-form commands go to the agent; the six quick bulk actions stay on the local planner.

The flow ​

  1. Say the intent — e.g. "ustaw cenę 100 wszystkim produktom bez ceny".
  2. The agent grounds it — it counts the matches first (never guesses) and may ask ONE clarifying question.
  3. It proposes a plan — a materialized diff, not a commit. The run shows awaiting approval.
  4. You review in the inbox — /agent/inbox: the exact before→after per object (∅ → 100), the scope (how many objects), the cost/tokens and the agent provenance badge. This is where you catch anything wrong — a malicious product description that tried to hijack the agent shows up here as an obvious sabotage, and you just reject it.
  5. Approve or reject — approve commits through the normal bulk path (provenance=agent); reject leaves the catalog untouched.

Undo ​

Committed a run and changed your mind? Settings → Agent → Historia → Cofnij tę operację within 24 h. Value and category changes revert cleanly. A schema change (new attribute) can only be undone while the attribute has no data — once someone filled it, undo is blocked and tells you why.

Privacy (what reaches the model) ​

Your Anthropic key is yours (BYOK) and never leaves the system in a response. The model sees your intent, the view context (object type, filter, selection counts — not full product data) and the results of the tools it runs (counts, codes, minimal projections). It never sees your key, passwords, API tokens or other tenants' data. See Settings → AI for the full list. Catalog writes happen only after your approval.

If the agent is unavailable ​

  • "agent wymaga klucza" — no Anthropic key is configured. Ask your operator (see the runbook).
  • A run refuses with a limit message — the tenant/your daily budget is spent; it resets at midnight UTC (or the operator can raise the cap).