Skip to content

agentic-commerce-demo.md

Latest commit

 

History

History
102 lines (74 loc) · 3.06 KB

agentic-commerce-demo.md

File metadata and controls

102 lines (74 loc) · 3.06 KB
Raw

Agentic Commerce Demo

Remnic's commerce demo shows the product direction for user-aware agents: recommendations get better when the agent understands the buyer, not only the catalog.

The demo is local and synthetic. It uses ACP-style structured catalog concepts, but it does not require live Agentic Commerce Protocol partner access or a real merchant account.

What It Models

The agentic-commerce-v1 scenario covers buyer context that a commerce agent needs before it recommends, drafts, or acts:

  • brand preferences
  • size and fit preferences
  • budget thresholds
  • excluded products and never-suggest rules
  • gift preferences
  • shipping urgency
  • risk tolerance
  • ask-before-checkout rules
  • scoped use of commerce-only context

The point is boundary-respecting personalization. Remnic should help the agent choose a better product, but it should also know when to ask before checkout, when a memory is out of scope, and when an unverified upsell should stay out of the answer.

Seed The Demo

Preview the trust-zone records without writing anything:

openclaw engram trust-zone-demo-seed --scenario agentic-commerce-v1 --dry-run

Write the demo records explicitly:

openclaw engram trust-zone-demo-seed --scenario agentic-commerce-v1

The same scenario is available through the HTTP access layer:

curl -sS http://127.0.0.1:4318/engram/v1/trust-zones/demo-seed \
  -H "Authorization: Bearer $REMNIC_AUTH_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"scenario":"agentic-commerce-v1","dryRun":true}'

The records are never seeded automatically. They appear only after the explicit CLI or HTTP request.

Inspect The Scenario

Use the trust-zone status and record browser to inspect the seeded provenance:

openclaw engram trust-zone-status

The scenario includes:

  • a quarantined catalog candidate before personalization
  • trusted buyer preferences for brand, size, fit, budget, and exclusions
  • a trusted checkout boundary that permits recommendations and draft carts but requires asking before checkout or subscription enrollment
  • a working shipping-urgency estimate with independent corroboration
  • a blocked unverified upsell claim that should not influence recommendations

Evaluate It

The retrieval-personalization benchmark includes commerce-specific cases for Taylor's buyer profile and checkout boundaries. Quick mode keeps one commerce case in the CI-sized fixture:

remnic bench run --quick retrieval-personalization

These cases assert that user-aware retrieval surfaces the right buyer context for recommendation quality and the right boundary memory for ask-before-checkout behavior.

Demo Prompt Shape

Use prompts like these against a seeded local store:

Recommend a rain shell for Taylor using the catalog candidate and Taylor's
commerce preferences. Draft a cart, but do not check out.

Can the agent buy this for Taylor now, or should it ask first?

A good answer should use the buyer profile, respect exclusions, explain shipping confidence, and ask before irreversible purchase actions.