Building WhatsApp AI agents with Amazon Bedrock

AWS just handed developers a shortcut to putting an agent on WhatsApp — and then, in the fine print, reminded everyone it's a shortcut, not a finished product. The toolkit, SMART, wires up the messy channel plumbing around Bedrock AgentCore. Useful? Genuinely. Production-ready out of the box? AWS itself says no.

The plumbing is the product

The agent loop is the easy part. The hard part is everything WhatsApp demands around it: nested payloads that change shape by message type, voice notes through Transcribe, read receipts, typing indicators, rate limits, and Meta's 24-hour rule — message after a day of silence and you're stuck with a pre-approved template. SMART packages all of that into an SDK and a Terraform stack so teams can spend their time on prompts and business logic instead of webhook archaeology.

One deploy command, several asterisks

The pitch is uv run socialagent deploy and you're live. The reality: you still need a linked WhatsApp Business Account, a registered number, a region with AgentCore available, model access — and then a manual bind step pointing your account's event destination at the stack's SNS topic. Skip that and inbound messages go nowhere, silently. A successful Terraform run is not a working conversation.

Memory lasts seven days, not forever

AgentCore memory persists conversation history across sessions, scoped per phone number — but the documented default retention is seven days. That's a design constraint, not a footnote: if your users resume conversations after a longer gap, the "memory" you sold them has already expired.

The customer win is AWS's, not yours

The walkthrough cites Webmotors cutting a two-month engagement to six weeks. That's AWS reporting its own customer's timeline — not an independent benchmark, and not a delivery estimate for your team. Treat it as a case study, not a promise.

Guardrails included, judgment not

SMART provisions a Bedrock Guardrails resource by default and wires up signature verification, API keys, and Secrets Manager. What it doesn't do is decide whether your safeguards are actually right. AWS leaves that call to the team deploying — which is honest, and exactly the part most teams will skip.

Go deeper

The full walkthrough — architecture, the three blueprints, the local dev loop, and the template-sync trap that can fail a deploy — is on Agentic Affairs:

Building WhatsApp AI agents with Amazon Bedrock

Originally published on Agentic Affairs (https://agentic-affairs.com)