Writing
Building reliable integrations
Webhooks lie, clocks disagree, and identity is a negotiation. Reliability is idempotency plus evidence.
- External systemsCRM, Slack, APIs
- IngestionWebhooks & fetches
- ProcessingIdentity, idempotency
- PersistenceEvidence before interpretation
- AIContext in, claims out
- DownstreamActions with retry policy
Default assumptions that will hurt you
- Exactly-once delivery.
- Stable IDs across systems.
- A meeting, a contact, and a Slack user being the same person on the first try.
- That a 200 from your webhook handler means the downstream write succeeded.
What “reliable” means here
- Persist the payload before you interpret it.
- Process with a key you can retry safely.
- Classify failures: auth, contract, downstream, poison message.
- Make the human-visible action (CRM note, Slack message) a job with its own policy.
This is the unglamorous half of production AI. Skip it and the model will confidently narrate a world your CRM does not contain.