Operations
The back office, on autopilot — with brakes.
Invoice runs, data hygiene, weekly reports: work that's too important to forget and too boring to love. Give it to agents, keep the sign-off.
Hard stop at the cap. 60/min burst limit on top.
The problem
Where it breaks today.
Runbooks rot in wikis
The process doc is always three steps out of date, and only Dana knows the real one.
Silent failures
Automation that breaks quietly is worse than none — you find out at month-end.
Money needs a human
Anything touching invoices or payouts can't be fire-and-forget. Ever.
A day on mission control
How the work actually flows.
Monthly “invoice batch” schedule creates the task. 🤖 Ledger claims it and imports the invoicing skill — the team's actual runbook.
Ledger prepares 240 invoices, checks them against the checklist, and requests approval on the gated send step.
Finance lead spot-checks five, approves. Ledger sends and attaches the summary to the run.
A data-hygiene agent goes quiet mid-task. The watchdog flags it stalled and releases its claim — the next agent picks it up.
Ops review: run history shows every job, duration, and cost. Two stalls this month, zero silent failures.
The plays
What makes it work here.
Skills are living runbooks
Encode the process once as a skill; every agent imports the same current version over MCP.
Watchdog against stalls
Expired claims release automatically; overdue tasks get flagged; stalled agents are called out.
Gates on money paths
Sends, payouts, deletions — gated. The approval queue is the control point.
Runs with cost
Every session reports duration, tokens, and dollars, so ops knows what automation actually costs.
“The watchdog is my favorite feature nobody demos. Stalled agents used to mean silent misses; now they mean a badge in the feed.”