An autonomous Meta Ads agent. No human in the loop.

A direct-response advertiser was running Meta Bid Cap campaigns at six-figure-monthly spend. Bid Cap requires hands-on operator attention — you set the bid, watch what Meta does with it, and adjust. If the bid is too low, the auction never unlocks and the ad set sits with zero delivery. If the bid is too high, you overpay. If ROAS drifts down, you have to react fast or burn budget. The advertiser wanted to scale without scaling the operator’s hours, and Meta’s built-in auto-bidder didn’t give them the discipline they wanted around a Real Net ROAS target (which had to factor real paid revenue and exclude certain product upsells).

We built an autonomous ad ops agent — a Python engine that runs every 30 minutes via cron, reads Meta ad performance and Whop sales from a Supabase warehouse, and runs each ad set through a three-phase state machine. Discovery raises the bid by $1 every 6 hours until the auction unlocks. Scaling grows the daily budget by 20% every 4 hours while Real Net ROAS holds above target. Corrective runs an escalating pullback ladder when ROAS drops — lower bid, then reduce budget, then pause — each step gated by three consecutive bad cycles. An emergency check runs every cycle and pauses immediately on spend-without-sales runaway or ROAS collapse. A fatigue monitor clones ad sets when frequency spikes or CTR craters. Six n8n workflows handle alerts to Discord, daily and weekly digests to Gmail, Meta token rotation, and a 15-minute health heartbeat.

The agent has logged over 300 autonomous decisions and 190 alerts in production — every bid change, every budget step, every pause, with a full audit trail in Postgres. The advertiser stopped hand-tuning ad sets and started running the configuration: tune the target ROAS, the emergency floor, the bid ceiling, the cycle cadence — the agent runs the playbook between cycles.