The morning my agents flagged my pipeline was empty

The brief that stopped me cold

My AI agents send me a morning brief at 6 AM. Every day. Standard thing.

Tuesday's brief had one line that made me put the coffee down:

Sales pipeline: empty. No active leads. No proposals in flight.
Last outreach: 11 days ago.

I'd been head-down on a client build for two weeks. Not a single proposal sent. The admin agent noticed before I did.

That's the pitch — and the warning. If you build a system that tells you the truth, you better be ready for it to tell you the truth.

What's actually running

A few of the agents I had online that morning:

• admin-agent — pipeline rollup, today's priorities, evening digest
• outreach-agent — Upwork proposals, cold DMs, follow-ups
• client-agent — STATUS.md sync per client, blocker flags
• content-agent — topic research, script progress
• research-agent — weekly market scan
• A few Struvo-specific ones (lead router, market intel, ops lead)

Each one runs on a different cadence. The admin-agent is the aggregator — it reads what the others did and writes me one digest twice a day.

What it cost

Total API cost the day this happened: about $6.40. I checked because the agents had been running in the background for two weeks and I wanted to know if I was burning money on autopilot.

I wasn't. The system pays for itself the moment it catches a missed outreach window. One late proposal sent on day 11 instead of day 21 is worth more than a year of API fees.

What I'd do differently

Two things I changed after that brief:

1. Add a "loud mode" to admin-agent. Empty pipeline shouldn't be one line buried in a digest. It should be a Telegram ping at 6:01 AM with a direct link to the proposal queue.
2. Tie the outreach-agent's daily quota to the pipeline state. If pipeline is empty, the daily proposal target jumps from 5 to 15 until it isn't.

Both changes took ~30 minutes to wire up. The agent fleet is configurable that way — you write a markdown spec, the system rebuilds itself.

What you can steal

The pattern is just: aggregator agent reads what the worker agents produced, writes one digest, names the thing that's broken.

You don't need a 15-agent fleet to do this. One Claude Code session with three skills hooked up and a cron job will get you 80% of the way there. The trick is not the agents — it's deciding what counts as "broken" before you build anything else.

The cost: about $5/day in API calls. The payoff: you find out about the empty pipeline on day 11, not day 21.