Three agents shipped meaningful work this week. Here is the work.

Most write-ups about AI in marketing are about the demo. This is about the output. One week of work from three agents on our team, what they shipped, what it cost, and what they got wrong.

The LinkedIn specialist

Seven post drafts, all on-pillar. Three published as-written. Two needed one round of edits. Two were cut for repeating an angle from the previous week.

Cost: about $0.40 in model spend across the seven drafts. Operator review time averaged 90 seconds per post. The same seven posts written by a freelance copywriter would have cost roughly $560 and three days of back-and-forth.

The numbers only work because the agent reads the same brief a human would: pillar, voice, hook, key points, CTA, sources. Thin briefs produce thin posts regardless of model size. We learned this the expensive way.

The calendar agent

It built a 30-day publishing schedule across five platforms, respected per-platform cadence, and moved three posts off Saturday after noticing the engagement curve flatlines on weekends in our category.

The interesting part is not the speed. It is that scheduling stopped being a context switch. The operator approved drafts. The slots filled themselves. The calendar review took four minutes on Friday morning instead of forty.

The manager agent

It paused twice. Once on whether to take a competitor angle on an X thread. Once on whether to rotate a Pinterest pillar this month. Both pauses arrived as one-line questions in Slack with a link to the draft.

This is the pattern that makes the whole stack viable: an agent that pauses on ambiguity instead of guessing. The operator answers when convenient. Production resumes without a meeting. The thing nobody warns you about with autonomous agents is that confident guessing on edge cases is what destroys brand trust over time. A pause is a feature.

What we got wrong

A duplicate draft for Instagram. Two agents pulled from briefs that pointed at the same source. We caught it before publish but after compose, which means we paid for the second draft. The fix was a one-line check at compose time. Cheap to add, embarrassing to need.

A hero image that rendered with the wrong section label on one slide. Operator caught it visually. The fix took ten minutes; the lesson is that visual review is still load-bearing even when generation is automated.

What this is not

Not a benchmark. Not a case study. Not a launch announcement. Just a log of one week of work, with the prices and the misses included. We will publish another one next week. The point is not that the numbers are impressive. The point is that they are auditable.