14 AI agents screened 120,000 influencers, ran the campaign and reported a 2.7% CTR lift

A post on dev.to by MiningLamp documents a multi-agent system that ran a complete influencer marketing campaign, from discovering creators to settling payments, with no human execution steps in between. Fourteen specialized agents screened roughly 120,000 influencer profiles down to 30, generated per-creator strategy and content, bought media programmatically, and reported a 2.7% improvement in natural click-through rate after a one-month trial.

How the pipeline is structured

The system is built as a hierarchy of narrow, specialized agents rather than one monolithic model. According to the post, each agent owns a single task and passes structured output to the next.

Data collection comes first. A footfall insight agent connects to the Lingxi and Juguang backends through authorized API integrations, which the post stresses is not scraping, and continuously pulls audience demographics, engagement patterns and content performance history. Because ingestion is continuous rather than a snapshot, the selection stage sees trajectory: a creator whose engagement has declined over 90 days is scored differently from one on an upswing, even when their current numbers look identical.

Selection is the headline stage. The agents evaluated about 120,000 profiles and kept 30, a reduction the post puts at 99.975%. The criteria go beyond follower count to include content style alignment, audience overlap with the target demographic, historical conversion rates and pricing efficiency. The author's argument is that the advantage over a human team is consistency: a junior marketer might manually review a few hundred profiles a day with fading attention, while the agent applies the same multi-dimensional scoring to every profile it sees.

Strategy, content, buying and settlement

Each of the 30 selected influencers received a customized strategy draft, described in the post as 60-point drafts, tuned to their format, audience segment, best posting times and the call-to-action styles that have historically converted for similar creators. The author credits this personalization stage alone with a 30% efficiency improvement over the manual process.

The content layer combines text-to-text generation, text-to-image creation and video editing in one pipeline, so visual assets are produced from the same brief context as the copy. The post claims content analysis that previously took a full session now saves three hours per round.

Media buying is fully automated: the agents connect to the Juguang API and purchase traffic directly, then reallocate budget in real time as conversion data flows back through the pipeline. Underperforming placements are throttled and strong performers receive more spend continuously rather than in morning review meetings. At the end of the chain, settlement is performance-based, with the client charged by sales leads and attribution tracked from impression through click, lead and payment inside the pipeline.

The reported results

After the one-month trial, the post lists a 2.7% improvement in natural click-through rate versus comparable campaigns, a 30% efficiency gain across the workflow, three hours saved per content analysis session, 14 agents spanning selection to settlement, and zero manual execution steps. The author notes that while 2.7% may look small, natural CTR benchmarks in influencer campaigns sit in the low single digits, and the lift came from better influencer-content matching rather than higher spend.

Why 14 agents instead of one

The post argues that a single mega-agent would have to hold every data source, API and optimization criterion in context simultaneously. The 14-agent design keeps each agent narrow, so the selection agent knows nothing about media buying, and lets them communicate through structured handoffs. The author likens the setup to a human team of strategist, media buyer, creative director and data analyst who pass work products between roles instead of doing each other's jobs.

The platform underneath

The pipeline runs on OCTO, an open-source, Apache 2.0-licensed, self-hosted workspace for human-agent collaboration. Its core concepts are agents configured to inherit a user's authorization and preferences, workspaces for collaboration, loops as work units with assignees, deliverables and acceptance criteria, and accumulated preference data from every review and rejection. The name stands for Open, Context, Taste, Orchestration. OCTO supports six collaboration modes: solo, roundtable, critic, pipeline, split and swarm. This campaign ran in pipeline mode, the sequential-handoff pattern.

Why it matters

Most public multi-agent demonstrations never leave the terminal window. This account stands out because it describes a production deployment with concrete, stage-by-stage numbers across a real commercial workflow, from data ingestion to payment. The caveats are real: the figures come from a single vendor-published post, the baseline of comparable campaigns is not defined precisely, and nothing has been independently verified. Even so, the architecture on display, with many narrow agents, authorized API integrations, structured handoffs and a closed optimization loop, is a concrete template for what multi-agent systems look like once they move past the demo stage.