Event-driven agent backbone

Agent workflows need to react to business events the moment they happen. Cron schedules and manual scripts can't do that — they poll, they miss, and when they fail mid-run there's no retry story, just a rerun and hope.

A shared event backbone: domain events (a quote created, an incident opened) are published to SQS, Lambda consumers fan out to the right agent workflow, and failures land in dead-letter queues with alerting instead of vanishing. The agentic selling product and the incident triage agent both run on it.

1. 01Domain event published
2. 02SQS queue (per consumer)
3. 03Lambda consumer invokes agent workflow
4. 04Result events emitted downstream
5. 05Failures → DLQ + alert

• MessagingAWS SQS with dead-letter queues
• ComputeLambda consumers
• Infra as codeTerraform
• ObservabilityDatadog monitoring + alerting

• One backbone powers multiple agentic products (selling, triage)
• Per-event retries and DLQs replaced the crash-and-rerun cycle
• Currently hardening: monitoring, alerting, and error-handling depth