A 24-agent operations fleet with one source of truth.
I designed and run the agent system that operates OnTriggerJoy and my own admin: a chat front door that delegates, standing routines for mail and money, an event-driven Slack listener with a watchdog, webhooks between two fleets, and one private GitHub repo that every agent reads first and writes back to with a verified commit.
16 of the 18 operational agents · own colour and symbol each · plus a 6-seat review council
OTJ agent fleet
Multi-agent automation for a small company: intake, routing, scheduled triage, event handling, records and approvals.
Project
- Owner
- OnTriggerJoy (my company)
- Role
- Architect and operator
- Scale
- 24 agents · 4 trigger types · 1 system of record
- Users
- Internal: me and the OTJ team. No external clients claimed.
What I built
- Routing: one dispatcher; every other agent has one job, explicit anti-jobs and its own write rules.
- System of record: a private GitHub repo with a MAP entryway, a write path with SHA verification and secrets kept as vault pointers.
- Event + schedule triggers: Slack Socket Mode listener, webhook wakes, weekday and weekly routines.
- Human in the loop: release tokens, approval-before-send, confirm-before-obey on inbound webhooks.
- Reliability: watchdog with self-repair, heartbeats, an atomic claim-and-dedupe drain, incident catch-up.
Stack
Links
Sources: bots/fleet/README.md and remits · docs/MAP.md · docs/ops/fleet-write-path.md · work/OnTriggerJoy/ops/slack.md + slack-drain-log.md. Private repo; architecture shown, contents not.


