# Monday checklist — LangChain on Amazon Bedrock

Compose the application in LangChain. Put models on Bedrock. Put production agent infrastructure on AgentCore when you outgrow a Lambda/API process.

## Day 1 — Model path only

- [ ] Enable Bedrock model access in the target region
- [ ] `ChatBedrockConverse` (or `init_chat_model(..., model_provider="bedrock_converse")`) with **explicit `max_tokens`**
- [ ] One prompt, one response, no tools — confirm IAM `bedrock:InvokeModel` on the inference profile
- [ ] If a single Converse call is the whole product, stop here — you do not need an agent framework

## Day 2 — Tools with a security boundary

- [ ] Define 1–3 tools as Python callables with narrow schemas (read-only first)
- [ ] Do **not** expose raw SQL or admin APIs to the model
- [ ] Prefer AgentCore Gateway (OpenAPI / Lambda / MCP) once tools leave the process
- [ ] Policy: LOG_ONLY then ENFORCE for any write tool

## Day 3 — Retrieval without dumping the estate

- [ ] Knowledge assistant: retrieve via Bedrock Knowledge Bases or a scoped search API, then generate
- [ ] Support: pass ticket + CRM fields as **application context**, not “the whole database”
- [ ] BI: controlled metric interface (pre-aggregated views), not ad-hoc warehouse access

## Day 4 — Agent loop vs graph

- [ ] `create_agent(model, tools, system_prompt)` for a bounded tool loop
- [ ] If you need explicit state, loops, HITL, or multi-agent routing, graduate to [LangGraph on AgentCore Runtime](https://www.factualminds.com/blog/langgraph-aws-bedrock-agentcore-2026/)
- [ ] If you need AWS-endorsed multi-agent primitives without a LangGraph investment, use [Harness / Strands](https://www.factualminds.com/blog/production-ai-agents-aws-agentcore-harness-strands-2026/)

## Day 5 — Host and observe

- [ ] Host on AgentCore Runtime (ARM64) **or** keep a thin API process if the loop is simple and you already operate it
- [ ] Attach Observability (OTEL → CloudWatch); golden eval suite before production traffic
- [ ] Browser / Code Interpreter **off** unless the turn requires them
- [ ] 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 unrestricted database tools
- [ ] No `max_tokens` left unset (quota reservation / throttling)
- [ ] No LangChain-vs-AgentCore “pick one” framing — they are different layers