Product · Loop engineering
ZeroCMO
Seven agent loops turn one X account into a measured feedback system with a human-gated write path.
- Result
- 3rd place
- Loop Engineering Hackathon SF
- Runtime
- 7 loops
- Scheduled and event-driven runs
- Roles
- 7 agents
- One role prompt per loop
- Write path
- 1 guarded
- Approve, edit, reject before X
01 Closed-loop system
Seven loops orbit one source of truth.
Each run reads shared state, produces one bounded output, and returns what the next run needs. Publishing closes a second loop through measured impact.
Draft → human gate → guarded write → X Post → metrics → Learning ↺
Score live X signals for the current goal.
- Reads
- Goal + live X signals
- Returns
- Ranked signals → memory
Mon · 06:00 · config $3.00 / 40t
Dollar and turn values are public v0.1 schedule configuration, not proof of runtime spend enforcement.
ZeroCMO was built for the Loop Engineering Hackathon in San Francisco. Seven scheduled or event-driven agent loops share one durable store: they sense X, watch competitors, evolve tone, draft, publish on approval, measure response, and feed a Learning into the next content run. The public v0.1 build placed third.
01 · Problem
Scheduled posting is a calendar, not a feedback system. The hard part is closing the path from a published post to measured response and then back into the next draft, without giving the model direct access to X credentials or permission to publish by itself.
02 · Architecture
Use seven narrow agent roles around a shared source-of-truth store. Content produces a draft, the operator approves, edits, or rejects it, and the sole credential-holding write server applies approval, duplicate, embargo, daily-cap, and kill-switch guards before posting. Impact measurements become Learnings that the next content run reads.
03 · What shipped
The public v0.1 hackathon build placed third and ships seven role prompts, seven configured loop schedules, a dashboard approval flow, and one guarded write path. Turn and spend ceilings are declared in schedule config; the case does not present them as demonstrated runtime enforcement or claim campaign lift.