14 agents filtered 120,000 influencers down to 30 and improved CTR by 2.7% — multi-agent in production

A developer deployed a production multi-agent system in which 14 specialized AI agents ran an entire influencer marketing campaign pipeline, from discovery through ad settlement, with no manual execution steps. The system scanned roughly 120,000 influencer profiles and narrowed them to 30 using multi-dimensional scoring, producing per-influencer strategy drafts that yielded a 30% efficiency gain over the manual process and a 2.7% CTR improvement. The agents also coordinated text, image, and video generation so visual assets matched the copy.

Most multi-agent demos stop at chatbots arguing with each other in a terminal. This one ran a real marketing campaign. 14 AI agents handled the entire pipeline from influencer discovery to ad settlement, with zero manual execution steps in between. I want to break down how this actually works, because the architecture is more interesting than the headline. A client needed a new product launch campaign. The traditional workflow looks like this: a human planner logs into multiple ad platforms, manually browses influencer profiles, builds spreadsheets, writes briefs, negotiates, tracks performance, adjusts spend. Each step involves a different tool, a different login, a different person. The bottleneck isn't creativity. It's coordination overhead. The system uses 14 specialized agents organized in a hierarchy. Not one mega-agent trying to do everything. Each agent owns a narrow task and passes structured output to the next. Here's the rough breakdown: A footfall insight agent connects to the Lingxi and Juguang backends via API. It pulls audience demographics, engagement patterns, content performance history across platforms. This isn't scraping. These are authorized platform integrations feeding structured data into the pipeline. The data ingestion is continuous. The agent doesn't just pull a snapshot. It monitors trends over time so that by the time the selection phase kicks in, the system has temporal context on each influencer's trajectory. An influencer whose engagement is trending down over 90 days gets treated differently from one on an upswing, even if their current numbers look identical. This is where it gets interesting. The system scanned approximately 120,000 influencer profiles and narrowed them down to 30. The selection criteria aren't just follower count. The agents evaluate content style alignment, audience overlap with the target demographic, historical conversion rates, and pricing efficiency. 120,000 to 30. That's a 99.975% reduction. A human team doing this manually might review a few hundred profiles in a day if they're fast. And even then, they're mostly eyeballing it. The agent applies consistent multi-dimensional scoring across every single profile. Think about what this means in practice. A junior marketer assigned to influencer research will open a platform, scroll through profiles, check a few metrics, maybe drop promising ones into a spreadsheet. After three hours they've got 40 candidates and their eyes are glazing over. The agent evaluates all 120k with the same rigor it applied to the first one. Once the 30 influencers are selected, the system produces 60-point strategy drafts. Not one generic brief. Each influencer gets a customized content deck based on their individual style and audience. If an influencer does comedic short-form video, they get a brief designed for that format. If another does long-form product reviews, different brief entirely. This personalization goes deeper than format. The strategy agent analyzes what topics resonate with each influencer's specific audience segment, what posting times correlate with their best engagement, what call-to-action styles have historically converted for similar creators. The output is a genuine per-person playbook. This alone resulted in a 30% efficiency improvement over the manual process. The content layer handles text-to-text generation, text-to-image creation, and video editing. All integrated into the same pipeline. The agents generate draft content that matches each influencer's voice and format preferences. This is where the multi-modal capability matters. A single agent doing text generation is table stakes in 2026. But having text, image, and video agents coordinated through the same pipeline means the visual assets actually match the copy. The hero image for a post was designed with the same brief context as the caption text. That consistency is hard to achieve even with human teams where the copywriter and designer work from the same brief but interpret it differently. Content analysis that used to take a full session now saves 3 hours per round. Here's the part that surprised me. The agents connect directly to the Juguang API and auto-purchase traffic. They optimize ad placement in real-time based on performance data flowing back through the pipeline. No human sitting there adjusting bids at 2 AM because a post went viral at midnight. The feedback loop is tight. Performance data from live campaigns feeds back into the optimization agent, which adjusts spend allocation across the 30 influencers based on real conversion data, not projections. Underperforming placements get throttled. High-performers get more budget. This happens continuously, not in morning review meetings. The system charges by sales leads. Performance-based pricing with automated tracking through to settlement. The attribution chain from impression to click to lead to payment is tracked end-to-end within the agent pipeline. After a 1-month trial run: 2.7% CTR improvement might sound small if you're not in marketing. In influencer campaigns, where natural CTR benchmarks hover in the low single digits, that's a meaningful lift. And it came from better influencer-content matching, not from spending more. You could theoretically build one giant agent that handles everything. But the coordination problem is real. Each stage has different data sources, different APIs, different optimization criteria. A monolithic agent would need to hold all of that context simultaneously. The 14-agent approach lets each agent specialize. The influencer selection agent doesn't need to know anything about media buying. The content generation agent doesn't care about settlement logic. They communicate through structured handoffs. This maps well to how human marketing teams actually work. You have a strategist, a media buyer, a creative director, a data analyst. They don't all do each other's jobs. They pass work products between roles. This pipeline runs on OCTO https://github.com/Mininglamp-OSS , an open-source workplace designed for humans and AI agents to collaborate. Apache 2.0 licensed, self-hosted. OCTO has four core concepts worth understanding: Agents are digital workforce doubles. They inherit your authorization and taste preferences. Not generic chatbots. They're configured to act within your specific context and permissions. Workspaces are where collaboration happens. Loops are work units. Each loop has an assignee, deliverables, and acceptance criteria. Think of them as the atomic unit of work that flows from conversation to delivery. Preferences accumulate over time. Every review, every rejection, every approval becomes experience data. The agents get better the more you work with them. The name OCTO itself encodes the design philosophy: Open, Context, Taste, Orchestration. For the orchestration piece, OCTO supports 6 collaboration modes: The marketing pipeline primarily uses Pipeline mode with some Split for parallel content generation. The repos are on GitHub: Self-hosted means your data stays on your infrastructure. For marketing campaigns involving client data and influencer analytics, that matters more than most people realize. The marketing use case is compelling because it's boring. Not boring in a bad way. Boring in the sense that it's a real business workflow with real money attached to it. Nobody built this to win a hackathon. They built it because the manual version of this process is slow and expensive. The 14-agent architecture is also interesting because it's not trying to be AGI. Each agent is narrow and good at one thing. The intelligence comes from the orchestration layer connecting them, not from any single agent being superhuman. Multi-agent systems that actually run in production, on real campaigns, with measurable business outcomes. That's the part worth paying attention to. After watching this pipeline run for a month, the takeaway that stuck with me is simple. The gains didn't come from having a smarter model. They came from having a smarter way to decompose work. Every time we tried to make a single agent handle more stages, performance degraded. Context windows filled up with irrelevant information from other stages. The influencer selection agent made worse choices when it also had to think about media buying constraints. Splitting responsibilities and connecting them through clean interfaces worked better every time. There's a parallel to microservices architecture here. Monoliths work until they don't. The moment your problem has enough independent dimensions, decomposition wins. Marketing campaigns have at least six independent dimensions: data, selection, strategy, content, distribution, and settlement. The 14-agent count isn't magic. Some stages use multiple agents internally. The selection stage alone has agents for audience analysis, style matching, and pricing optimization that feed into a ranking aggregator. The point is that granularity at the agent level mapped naturally to granularity at the task level. If you've deployed multi-agent systems in production, what coordination patterns worked for you? Pipeline seems natural for linear workflows, but I'd be interested in hearing about cases where Swarm or Roundtable modes proved more effective.