Mental Model

Overview

The mail path is easiest to hold as five layers. AI steps sit in Decide and Act thank-you; everything else is infrastructure or people. Anthropic names from Building effective agents (PDF).

Applied workflows

• Routing. Inbound System 1 ranks clean mail into an allowlisted category slug (gpt-4o-mini or Typesafe Jev); thank-you framing uses that label.
• Parallelization. Forward and thank-you are independent send concerns (separate retries / idempotency). Diagrams show Processing Q and Thank-you Q in parallel; the deployed worker may still run both in one outbound Lambda with per-step skip.
• Not used here (yet). Prompt Chaining, Orchestrator-Worker, Evaluator-Optimizer.

Layers

• 1 Ingest. SES receive stores raw MIME under ses-raw/; spam/virus verdicts arrive with the receipt.
• 2 Decide. Inbound Mail claims the message, ranks category (Routing), stores under inbound/, marks queued.
• 3 Dispatch. SQS carries the job id (and metadata) to the send worker — claim-check style, not the full MIME.
• 4 Act. Forward to FORWARD_TO and thank-you to the sender (Parallelization of concerns); each step records DynamoDB send ids.
• 5 Observe. mail-events (bounce/complaint via SNS), CloudWatch metrics, DLQ alarms.

Layer map

Layer stack for one inbound message — Routing on Decide, Parallelization across Act. Diagram key: ✦ AI agent · second line = Anthropic workflow · purple = actor · white = service.

---
config:
  theme: default
  themeVariables:
    background: "#ffffff"
    fontSize: "14px"
---
flowchart TB
    person[Person]
    ingest["1 Ingest
SES plus S3 raw"]
    decide["✦ 2 Decide
Routing"]
    dispatch["3 Dispatch
SQS outbound"]
    actFwd["4a Act forward
SES to FORWARD_TO"]
    actThanks["✦ 4b Act thank-you
Parallelization"]
    observe["5 Observe
mail-events CloudWatch"]
    inbox[Personal inbox]
    reply[Thank-you to sender]

    person -->|email hello@| ingest --> decide
    decide -->|action queued| dispatch
    dispatch --> actFwd --> inbox
    dispatch --> actThanks --> reply
    reply --> person
    ingest -.-> observe

    classDef agent fill:#e8f0ee,stroke:#0f3d36,stroke-width:2px,color:#0d0d0d
    classDef service fill:#fafaf8,stroke:#c8c8c2,stroke-width:1px,color:#0d0d0d
    classDef actor fill:#f3eef8,stroke:#6b4f9a,stroke-width:2px,color:#0d0d0d
    class decide,actThanks agent
    class ingest,dispatch,actFwd,observe,inbox,reply service
    class person actor

Read layers top to bottom; drill into agent tabs for verdicts, queues, and retries.