Designing Permissions for an AI Agent on a Real WhatsApp Account
Apart_Buy5500 · reddit · 2026-07-15
The author connected an AI agent to their personal WhatsApp number via a custom bridge, rather than using the official Business API. The key risk here is that if the agent says the wrong thing, it sends directly to real-life contacts, damaging relationships rather than just causing a software error.
To mitigate this, they heavily focused on permission design:
- Implemented a contact/group whitelist, denying access by default
- Most chats are restricted to read and summarize only, with no sending capabilities
- Media permissions (like images) are controlled separately from text permissions
- All messages, candidate replies, and actual sends are logged for playback and debugging
- Included a physical kill switch independent of the agent itself
They noted that the most valuable daily use case isn't "automating customer replies," but rather having the agent monitor a high-volume group, filter out a few actionable messages daily, and recap chat context when they are busy. The core takeaway: for an agent on a real number, the true failure mode isn't "giving a wrong answer," but rather "saying something technically correct in a context where it shouldn't have spoken at all."
More from coding & agent
- A GLP1R variant may explain stronger Ozempic weight loss, and the team built an agent workflow — julia_kiseleva · 2026-07-21
- A Claude-coded Chrome extension shames you with a private jet when you open YouTube — alex_verem · 2026-07-21
- A curated TTS list for voice agents tracks latency, cancellation, and evals — mahimairaja · 2026-07-21
- Harness engineering is emerging as the execution layer for reliable AI agents — Pavan_Belagatti · 2026-07-21
- DevFest Lisbon keynote will cover Google AI Studio’s latest vibe coding and agentic AI features — gerardsans · 2026-07-21
- Daniel Hanchen’s 2-hour workshop covers open models, reward hacking and RL — danielhanchen · 2026-07-21