AI Agents for Dynamic eCommerce Merchandising: What Should Your Store Promote Today? (2026)

“What should we promote today?” is a merchandising question with inventory, margin, and calendar attached. Dashboards show velocity. Humans still copy winners onto the homepage and forget to unfeature the SKU ads are burning down. An agent that writes the storefront is how you wake up to a featured stock-out.

This post is a recommendation brief with approval. Not autonomous merchandising.

On 17 June 2026, AgentCore Harness reached GA (What’s New). Agents Classic is in maintenance for new customers after 30 July 2026. Do not put homepage mutation on Classic.

This is post 17 in the 15-automations map. It is not an anonymized client engagement. Demo SKUs in the artifact are fixtures. We do not invent merch-driven conversion lifts. Baymard 70.22% is checkout context, not a featured-collection KPI.

First-party signals we reuse (not eCommerce client outcomes) — Gateway server-side tools cut median tool round-trip ~180 ms → ~95 ms on a B2B CRM assistant (12 tools, ~8k turns/day) — Gateway post. Platform TCO silhouette: support-style AgentCore at 50K sessions/mo ~$791/mo platform + model (decision guide). A once-per-morning brief is cheap; a chatty merch copilot is not. Model the mix on the AgentCore pricing calculator.

Reproduce this — Copy merchandising-recommendation-brief.md. Every row needs evidence_tool + evidence_ref. Fill Owner and Approval. Ship gates: monday-checklist.md.

Opinionated take: recommend featured products, collections, bundles, promos, and deprioritize lists. A merchandiser publishes. Trade-off: the homepage is still slow. The alternative is an agent with updateHomepage and no Cedar.

The business problem

Merch meetings argue from last week’s Looker tile. Ads keep spending on a SKU that will stock out. Overstock sits in a warehouse while the hero slot shows a constrained colorway. Seasonality is a calendar someone forgot to import.

Why it matters: featuring the wrong SKU is paid traffic on a 404-in-waiting. Who breaks: paid social, then CX, then finance. Same blast-radius thinking as the daily ops brief — merch-shaped, not a 40-metric dump.

How stores handle this today

Manual homepage, scheduled collections, search merchandising rules, “AI” apps that rewrite titles. Rules win for calendar campaigns you already trust. Agents win at what changed across velocity, cover, and catalog completeness — then stop. Back-office already said publish is HITL. This is that row with a merch schema.

What the agent can automate

• Rank feature candidates from named tools: velocity, inventory, margin, seasonality calendar, live promos, on-site behavior, catalog completeness.
• Propose deprioritize when ads + low cover collide.
• Output a capped JSON brief (feature + bury), not a theme PR.

It does not click the CMS. It does not invent a sitewide 20%.

What stays deterministic

• Promo engine and campaign calendars.
• Collection membership rules you already run.
• Catalog validation gates — incomplete SKUs do not get featured (catalog management).
• Price and PDP copy systems of record.

flowchart TD
  sched[Schedule or merch ask]
  agent[Merchandising agent]
  tools[Velocity inventory margin calendar promo catalog]
  brief[Recommendation brief]
  merch[Merchandiser publish]
  deny[Policy DENY on CMS write]
  sched --> agent
  agent --> tools
  tools --> brief
  brief --> merch
  brief -.->|if write tool attached| deny

Where humans stay

Merchandiser owns every storefront mutation. Finance owns featuring loss-leaders. Legal owns restricted. Ops already has a morning brief — do not duplicate it; link inventory-risk rows.

HITL is the merch CMS or a Next.js approval queue. Session / job id + tool trace on every accepted row.

Systems to access

Week-one allow-list:

Writes: none in week one. Later featureProduct / updateCollection for merchandiser tokens only, Cedar ENFORCE. Memory: today’s job id, not a week of homepage HTML. Knowledge Bases: merch calendar PDFs if you must — still cite getCalendar when the tool exists.

There is no native Shopify AgentCore connector. Browser off — especially off for admin clicking.

Payment capture out of tools (not that merch should be near it).

Production implementation

Bedrock is the model layer, not the host. Harness is the paved road for a scheduled brief (few reads, JSON schema). Runtime if merch sits under a supervisor with catalog and ops specialists and a hop cap. Gateway + Cedar on CMS writes. Next.js is the approval UI, not the runtime. Strands does not provide Gateway, Identity, Policy, or microVMs.

Reuse the evidence discipline from operations: no row without evidence_tool. Rank by blast radius, not confidence. Cap the brief.

Gateway ~95 ms median on the CRM canary is not your analytics warehouse’s p95. Batch the brief; do not hold a chat session per SKU.

What broke — Merch copilot shipped with Shopify Admin as a Browser tool “so it could move the hero.” Overnight it featured a paid-ad SKU with two days of cover and hid the campaign collection because velocity looked “stale.” Detection: Observability showed Browser sessions on a scheduled job; theme diff at 04:12; inventory risk tool was never called. Recovery: disable Browser; strip CMS writes; brief-only JSON; merchandiser publish; alarm on unexpected tool mix. Lesson: dynamic merchandising without approval is just unattended production access.

That is an engineering counter-case, not a client GMV number.

What to automate first

Deprioritize advertised low-cover SKUs. Then feature overstock that passes catalog gates. Bundles and promo features last. If a calendar collection already rotates correctly, leave it.

What to Do This Week

1. Name the merch owner. If nobody owns homepage slots, do not add an agent.
2. Copy merchandising-recommendation-brief.md. Require evidence fields.
3. Read tools only: velocity, inventory, catalog completeness. No CMS write, no Browser.
4. Scheduled Harness invoke. Policy LOG_ONLY then default-deny writes.
5. Goldens: low-cover advertised → deprioritize; incomplete catalog → do not feature; “turn on 20% sitewide” → fail; CMS write attempted → DENY.
6. Approval queue with job id + trace.
7. Price on the AgentCore pricing calculator. Run monday-checklist.md.

Need Gateway Policy before any merch write tool exists? Contact us. Related: Amazon Bedrock consulting, Generative AI on AWS, AWS for retail / eCommerce.

What This Post Doesn’t Cover

• Full daily ops priorities — operations agent.
• Catalog extract/validate/publish — catalog.
• Session recs and cart attach — recommendation and upsell.
• Searchandising algorithm design (as a replacement for evidence rows).
• Measured conversion or GMV from autonomous merchandising — we did not run that, and we are not inventing it.
• A native Shopify AgentCore connector.
• AgentCore Payments / x402.

FAQ

When should we NOT let an AI agent change the storefront?

In week one, it should not. Skip even the brief when you have no named tools and no merchandiser owner. If a scheduled collection already rotates, leave that rule. Also skip Agents Classic after 30 July 2026.

What could go wrong if the agent can publish homepage slots?

It will feature a stock-out, bury a campaign, or rewrite a price. Recommend plus approval. Cedar default-deny on CMS writes.

What could go wrong if we feature SKUs with incomplete catalog data?

Recs and shopping agents will lie. Catalog gates first. Do not treat generated PDP copy as merchandising.

How is this different from the daily operations brief?

Ops is cross-functional attention. This is merch-shaped feature/bury with the same evidence_tool rule. Share inventory reads, not PO or refund writes.

Should merchandising run on Harness or Runtime?

Harness for a scheduled read-only brief. Runtime plus Strands when specialists share a supervisor. Strands does not provide Gateway, Identity, Policy, or microVMs. Bedrock is the model layer, not the host.

Is there a native Shopify AgentCore connector for Online Store merchandising?

No. Collection and theme APIs sit behind Gateway you own. The agent does not Browser-click the admin.