AI agent orchestration — 4 systems, 3 incidents, 5 collaboration rules
Summary in three lines
- AI agent teamwork depends on structure, not intelligence: role boundaries, gates, one ledger and clean handovers.
- With four execution systems each working hard on its own, the same post went out twice, a dead bot showed a green light and one website listed four different team sizes.
- Five rules followed: keep instructions in the chat channel, hand over via documents, track status in one ledger checked against real output, turn incidents into gates and have another system challenge key conclusions.
What we learned
- Do not just confirm a scheduled task is registered; confirm it can actually run. A default that blocks scheduled tasks on battery power once stopped jobs silently.
- Read every number from one master file and block deployment when pages disagree. Numbers edited by hand eventually drift again.
FAQ
How do you confirm a deleted post is really gone?
Do not rely on a 200 response, because a deleted post once returned 200 too. Presence in the RSS feed became the deciding check.
Are there limits still unsolved?
Yes. The measurement system is paused after a device change so only three run today, checkers sometimes flag healthy jobs as failed, and bot reports can misread causes.
This page is a summary BELLA AI wrote after reading the original. The full process and screenshots are in the original post.