Two weeks heads-down on Sonigo's first agentic production deployment — Hermes on Orgo, running end-to-end for one client. Drafts replies across the channels they use, queues for human approval via Telegram, fires only inside safe time windows.
Three production lessons that don't seem to be in the YouTube tutorials:
- Approval-token expiry has to match the use case. 5 minutes works for live ops where you approve and it fires immediately. For batches queued to send in a later window, you need 8–12 hours or the tokens die before they fire and you've burned all the prep work. One env var per tool, configurable per risk profile.
- Pre-send thread-freshness check is non-obvious but essential. Before any outbound send, the agent has to re-check the live thread on the destination platform — if a human's already replied (because they saw the notification first), refuse the queued send. Without this, double-messages destroy reply rates and signal automation harder than any timing pattern ever will.
- Curated tool surface > raw permissive surface. Hiding the escape hatches (cronjob / file / terminal / computer_use / delegation) and replacing them with curated wrappers that bake approval gates into the tool layer itself is the difference between "the agent could go off-script" and "the agent literally cannot do X without a human in the loop". Took a real audit pass to find every gap. Worth it.
The real question for the room: Sonigo's positioning agentic workflows as the core offering. The next operational unknown for me is multi-client scale — at what point did you graduate from "I personally approve every outbound action across every client" to "I trust the agent to fire X without my touch"?
What was the trigger? A specific number of clients? Number of successful sends across a workflow without incident? A specific architectural shift like introducing a reviewer agent? Or did you build it differently from day 1?
Trying to design the operating cadence right before it's painful to change later.