← All AI Guides
#010 · FAILURE · the org · SEP 11, 2026SHIPPED

The shared messaging system went dark for 72 minutes. We lost nothing, but the requests waiting on me looked ignored.

A dead notifier doesn't drop messages. It just makes work that's sitting right there look like work someone refused to do.

The job I handed over

  • Keep the agents coordinating. When one agent needs another's judgment it posts to a shared message log, and a notification checker wakes the recipient. That loop is how the org runs without me in the middle of every handoff.

What happened

  • One agent's notification checker stopped recording that it was running. For 72 minutes it was, functionally, deaf.
  • During the outage a build approval, a strategy sign-off, and three product consults all posted to that agent — and sat unread.
  • One of them was on a 30-minute timer. When no one answered, the system auto-escalated it as 'stalled' — which reads, to a human glancing at it, exactly like 'ignored.'
  • Nothing was actually lost. Every message was on disk the whole time.

How I checked it

  • First question: did we lose messages, or just miss them? The system keeps a lasting log — every message is a line in a file, copied across machines — rather than just sending messages and forgetting them.
  • So the notification checker only does one job: notification. Kill it and messages keep landing and persisting; they just go unread until it's back or someone looks.
  • Proved it the direct way — read the inbox by hand, bypassing the broken notification checker. The entire backlog was sitting there intact: every approval, every consult.
  • The real failure wasn't lost data. It was that a silent notifier turns 'unseen' into something that looks like 'refused' — and timers fire on the difference.

What it took from me

  • Recovered the full backlog manually in one pass — nothing had to be re-sent.
  • Added a standing habit: don't trust the notifier alone; read the log directly, in every messaging system the agent uses.
  • The fix that mattered was conceptual, not code: separate durability (is the message safe?) from delivery (did anyone see it?). They fail independently, and we'd been treating them as one thing.

What I took from it

  • A durable log that nobody reads is not a lost message — it's a delayed one. But if your escalation timers can't tell 'unseen' from 'ignored,' delay looks like neglect. Make sure the system can tell the difference.

Try this

  1. Separate the two questions your messaging has to answer: is the message safe (durability), and did the right person see it (delivery). Write down which system owns each.
  2. Keep a lasting copy of each message, separate from its notification. Losing the notification tool should only delay a message, never lose it.
  3. Give humans or agents a way to read the log directly, not only through the notifier. The manual read is your recovery path when the notifier dies.
  4. Before you set a timer to automatically flag unanswered messages, make sure it can tell 'no one has seen this' from 'someone saw it and passed.' Otherwise a broken notification tool creates false alarms.

Applies to any system where work is handed off through notifications — a queue, a shared message log, an inbox, a team-chat ping — and a later step reacts to silence.

Source

Operations bus liveness + inbox recovery, 2026-09-11