Agent fleet
More than 30 scheduled agents run my businesses, with written rules on what each may do alone.
How it works
Scheduled runs are held during the day and released later, skipping any that a later run already covers.
Claude agentsClaude Code agents start on a schedule, with no one at the keyboard.
Hermes agentsA second platform, Hermes, runs its own agents. They sit in the same list as the Claude ones, so one view shows the whole fleet.
Each agent judges an action by who it reaches and whether it can be undone. Private and reversible, it just does it. Public, it shows me what changed. Permanent or costing money, it asks first.
Wake on a messageA webhook on Vercel turns each new Slack message into a task for a responder agent, so a message doesn't wait for a schedule.
Clear names in SlackEach platform posts in Slack under its own name, so I always know which system said what.
A script compares the live schedules with the written record and flags any agent still running from instructions that were deleted.
What went into it
- The guiding rule: silence is not success. "Nothing needed doing" and "I am broken" must look different.
- The health check's error count sits on my operating dashboard, where it should read zero.
- What an agent may do alone depends on the risk of the action, not its subject.
Built with
Claude
Hermes Agent
Slack
Vercel