# Monday checklist — LangGraph on AgentCore Runtime

Keep the graph in application code. Put production operations on AgentCore. Do not host a greenfield LangGraph loop on ECS unless you already operate that fleet.

## Day 1 — Confirm Runtime is the right host

- [ ] You need explicit graph topology (conditional edges, loops, HITL) **or** you already have a LangGraph you will not rewrite this quarter
- [ ] If a single-domain config loop is enough, stop and use [AgentCore Harness](https://www.factualminds.com/blog/production-ai-agents-aws-agentcore-harness-strands-2026/) instead
- [ ] Pick a [supported AgentCore region](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/agentcore-regions.html)
- [ ] Confirm Bedrock model access for the inference profile you will pin

## Day 2 — Graph + Bedrock Converse

- [ ] Define a `TypedDict` / Pydantic **State** schema (messages + business fields — not “the whole prompt”)
- [ ] Nodes for classify / retrieve / tool / HITL / close (or your real topology)
- [ ] `init_chat_model(..., model_provider="bedrock_converse")` with **explicit `max_tokens`**
- [ ] Compile locally; no AWS required for a dry graph invoke with mocked tools

## Day 3 — AgentCore Runtime entrypoint

- [ ] Wrap `graph.invoke` in `@app.entrypoint` on `BedrockAgentCoreApp` (see [`langgraph_runtime_stub.py`](langgraph_runtime_stub.py))
- [ ] Build **ARM64** (Graviton) — x86 images will not start
- [ ] Health check on port 8080 per Runtime protocol
- [ ] Create Runtime + Runtime endpoint; wait until `READY`

## Day 4 — Memory and tools

- [ ] Attach `AgentCoreMemorySaver` with a real Memory ID; pass `thread_id` **and** `actor_id` on invoke
- [ ] Do **not** treat LangGraph checkpoint blobs as a substitute for AgentCore long-term Memory strategies
- [ ] Put enterprise writes behind **Gateway + Policy** (LOG_ONLY then ENFORCE) — graph edges are not an authorization boundary
- [ ] Browser / Code Interpreter **off** unless the turn requires them

## Day 5 — Observability and cost

- [ ] Confirm OTEL spans in CloudWatch for node/tool failures
- [ ] Golden eval suite before you trust “the graph works”
- [ ] Model platform + tokens on the [AgentCore pricing calculator](https://www.factualminds.com/tools/amazon-bedrock-agentcore-pricing-calculator/)

## Hard stops

- [ ] No Agents Classic for net-new builds (maintenance for new customers after Jul 30, 2026)
- [ ] No LangGraph-on-Harness (wrong product)
- [ ] No unrestricted database tools from the model
- [ ] No HITL skip on high-impact writes (refunds, IAM, production deploys)