# Monday checklist — Harness first, Strands when topology gets real

Ship a grounded agent this week without sequencing on AWS Context.

## Day 1 — Harness in non-prod

- [ ] Pick a supported AgentCore region
- [ ] Create IAM execution role for Harness (scoped Bedrock + `bedrock-agentcore:*` as needed)
- [ ] `agentcore create --name <agent> --model-provider bedrock` **or** `create-harness` via control plane
- [ ] Poll until status `READY`; invoke with a ≥33-char `runtimeSessionId`
- [ ] Confirm CloudWatch log group / spans appear for the harness

## Day 2 — One real tool via Gateway

- [ ] Wrap one internal read API (OpenAPI or MCP)
- [ ] Register as Gateway target; confirm Identity/credential brokering (agent never sees raw API keys)
- [ ] Policy: LOG_ONLY first if the tool can write; ENFORCE before prod
- [ ] Cap Browser / Code Interpreter — off unless the turn requires them

## Day 3 — Grounding that ships

- [ ] Create Managed Knowledge Base; connect one S3 prefix of markdown runbooks
- [ ] Attach KB as Gateway target; ask a multi-hop ops question
- [ ] Do **not** wait on AWS Context (coming soon)

## Day 4 — Eval baseline

- [ ] Define 10–20 golden tasks with expected tool calls / answers
- [ ] Enable Observability dashboards (latency, tool errors, cost attribution)
- [ ] Plan A/B only after you know what `task_completion_rate` means for your suite

## Day 5 — Decide graduation

- [ ] Still single-domain and config-enough? Stay on Harness
- [ ] Need Graph/Swarm/Agents-as-Tools or hop caps? Export to Strands → Runtime
- [ ] Model platform + tokens on the [AgentCore pricing calculator](https://www.factualminds.com/tools/amazon-bedrock-agentcore-pricing-calculator/)

## Hard stops

- [ ] No Browser-on-every-turn in prod
- [ ] No AWS Context in the critical path
- [ ] No Classic Agents for net-new builds (maintenance for new customers after Jul 30, 2026)