# Monday checklist — first eCommerce AI agent

Paste this into the RFC. Ship **one** customer-visible or ops-visible workflow this week. Do not start five agents in parallel.

## Day 1 — Pick the surface

- [ ] Write the business problem in one sentence (example: "WISMO tickets after carrier scan stalls")
- [ ] Name the systems the agent may **read** (OMS, carrier, catalog, helpdesk)
- [ ] Name the actions that need **human approval** (refund, cancel, inventory write, vendor PO)
- [ ] Choose Harness if ≤5 tools, one team, no hop caps — otherwise Runtime + Strands
- [ ] Confirm Agents Classic is **not** the net-new path (maintenance for new customers after 30 Jul 2026)

## Day 2 — Tool contract

- [ ] Draft OpenAPI (or MCP) for **read** tools only
- [ ] Attach Gateway; Policy mode `LOG_ONLY`
- [ ] Default-deny writes; no refund/cancel/inventory-adjust tools until Cedar is reviewed
- [ ] Browser / Code Interpreter **off** unless the turn needs them

## Day 3 — Eval suite

- [ ] 10 golden tasks with expected tool + evidence field
- [ ] 3 "must escalate" cases (legal language, chargeback, PII dump)
- [ ] Alarm on Policy DENY spikes and unexpected tool mix
- [ ] Model cost on the [AgentCore pricing calculator](https://www.factualminds.com/tools/amazon-bedrock-agentcore-pricing-calculator/)

## Day 4 — Human path

- [ ] HITL queue exists (helpdesk, Slack, or Next.js ops dashboard)
- [ ] Session id + tool trace attached to every escalation
- [ ] Shopper vs associate Identity claims tested (associate-only writes DENY for shopper tokens)

## Day 5 — One proof, one owner

- [ ] Non-prod invoke of **one** harness or Runtime
- [ ] Owner named for evals + cost alarm
- [ ] Book a review if you need Gateway Policy on refunds — [contact](https://www.factualminds.com/contact-us/)