One Effect program.
Agent + infrastructure.

The same agent, four ways to ship it

The AgentCore samples deploy with YAML, HCL, or a second TypeScript runtime that only describes infrastructure. Alchemy collapses the gap: resources are values in your Effect program, so a failed CreateAgentRuntime is a tagged error you handle — not a stack event you go hunting for.

Nine lessons, source-grounded and runnable

Build an Effect-docs agent and deploy it to Amazon Bedrock AgentCore — agent, tools, memory and observability — from a single typed Effect program with Alchemy.

Milestones: runs locally at 02–03 · first deploy at 05 · full stack by 08

1. 00Project setup & installing dependenciesToolchain (Node + pnpm), the pinned dependency set, workspace settings, tsconfig; verify with pnpm test.
2. 01What AgentCore is, and why Infrastructure-as-EffectsThe two-route runtime contract; the "one Effect program" pitch; CDK/TF/CFN vs Alchemy.
3. 02The agent as an Effect HttpApi/ping + /invocations on :8080; the Init vs Runtime phase split. runs locally
4. 03Talking to the model: a minimal BedrockModelA provider-agnostic LanguageModel over Bedrock Converse; stub ⇄ Bedrock as one line.
5. 04Containerize the agent and push to ECRThe container that serves the contract; declare the registry in the same program.
6. 05IAM role and the AgentCore RuntimeExecution role + the given Runtime resource; deploy + invoke for real. first deploy
7. 06Gateway: explainer search as an MCP toolsearchExplainers as an Effect Lambda, fronted by a Gateway MCP target.
8. 07Memory: learner progress across turnsPersist context across invocations; read at turn start, write at end.
9. 08Observability: tracing the agentEffect spans around the model and tools; OTEL → X-Ray / CloudWatch, connected across services. full stack
10. 10Appendix: authoring the AgentCore resourceHow the given Resource + Provider is built against the control-plane SDK.

Local-first, then real