Skip to content
Palawan Collective

044

Filed
6 September 2026
From
El Nido · Remote
Reading time
10 min

Deploying AI Agents in a Province with Unstable Internet

Our agents assume the connection will drop. Here’s the architecture that lets them queue, retry, and still hand every message to a human on time.

A bamboo hut in dense tropical greenery
Fig. — A bamboo hut in dense tropical greenery

The first agent we deployed assumed it was always online. It answered beautifully — until a squall knocked out the uplink for forty minutes and nine guest messages sat unanswered, with no record of what had been promised.

§Design for the drop

  • Every inbound message is written to a local queue first, then processed. If the uplink drops, nothing is lost.
  • Agents never confirm a booking or a payment on their own. They draft; a human approves from their phone with one tap.
  • Two uplinks: satellite as primary, a 4G router on a different network as failover. The switch is automatic and the agent doesn’t care which one it’s on.
  • When both are down, the caretaker has a printed fallback: prices, availability rules and the boatman’s number.
Field log — Squall, 14:02–14:51
  1. 14:02Satellite obstruction — heavy rain cell. Failover to 4G.
  2. 14:094G degraded. 3 messages queued.
  3. 14:44Uplink restored. Queue flushed: 9 replies sent, 2 flagged for human approval.
  4. 14:51Owner approved both from the back of a tricycle.

“An agent that can’t admit it’s offline is worse than no agent.”

§What the agents actually do

Mostly unglamorous work: answering “is there availability on the 14th?” at 11 p.m., collecting passport names for boat manifests, reminding guests that pickup time depends on the tide, chasing deposits politely. The value isn’t intelligence. It’s that nothing falls through the cracks while the humans are asleep, at sea or out of signal.

— D.Filed from El Nido · Remote