Egg & Geese
Inspiration
We watched brands burn money on influencer deals and ad placements that feel fake, get skipped, and convert nobody. Meanwhile, the comments that actually move product are the ones that sound like a real person who genuinely solved a problem. We wanted to build a system where marketing is not broadcasted at people but woven organically into the conversations they are already having. The idea was simple: what if a swarm of AI agents could find real pain-point discussions across social platforms and enter those conversations with authentic, product-accurate responses that actually help people, then learn from what works and double down on it?
What it does
Egg & Geese is a multi-agent, self-evolving social media marketing platform. A business describes their product to a conversational agent, and the system takes over from there. GLiNER-powered zero-shot extraction anchors the product's identity: pain points it solves, key benefits, active ingredients, target audience. A swarm of scout agents fans out across Twitter, Reddit, and Instagram using Yutori to find live discussions where people are complaining about the exact problem the product addresses. Reka Vision analyzes visual context from post images and media to give the agents richer situational awareness before they engage. The strategy layer, powered by Claude, then crafts responses that feel deeply human: a genuine recommendation, a casual mention, an empathetic reply that happens to reference the product. Senso serves as the product knowledge backbone, ensuring every generated comment stays factually grounded and never hallucinates product claims. OpenClaw handles the actual execution, posting comments, replies, and reposts across platforms through a dedicated gateway service. After engagement, the system tracks real performance metrics at regular intervals: impressions, likes, replies, reposts. All of this, every agent action, every strategy, every metric, gets stored as a knowledge graph in Neo4j. The learning agent analyzes what worked and what flopped at the entity level, identifying which pain points, tones, and strategies drive the most engagement, then feeds those insights back into the next cycle. Strategies that underperform get deprioritized. Approaches that blow up get amplified. The swarm evolves.
How we built it
The orchestrator runs on FastAPI with a six-stage agent pipeline: Intent, Scout, Vision, Strategy, Engagement, and Learning. Each stage receives a GLiNER-built campaign schema that keeps entity consistency across the entire flow. GLiNER 2, accessed through Fastino's hosted API, is the perceptual backbone of the system. It powers zero-shot named entity recognition at every stage: extracting product profiles, scoring scouted posts by entity overlap, validating generated comments for accuracy, and analyzing top-performing content to find winning entity patterns. Claude handles the reasoning and generation layer, crafting humanized comments and orchestrating strategic decisions. Yutori provides the web agent infrastructure for continuous social media monitoring and deep-dive research. Reka Vision processes images and video attached to scouted posts, giving the strategy layer visual context that text-only analysis would miss. Senso acts as a verified product knowledge base so the agents never drift from accurate product information. The execution gateway is a separate Node.js service built on Express that interfaces with platform APIs directly: twitter-api-v2 for Twitter, snoowrap for Reddit, and Puppeteer-based automation for Instagram. Neo4j stores the full knowledge graph of campaigns, products, strategies, engagements, and entity-level performance data, forming the memory that powers the self-improvement loop. PostgreSQL handles campaign metadata and relational state. The frontend is built with Next.js 15 and React 19, featuring real-time agent activity streaming over WebSocket, force-directed swarm visualizations, pipeline stage tracking, strategy leaderboards, and knowledge graph exploration. Everything is containerized with Docker Compose for local development, with a path to Render for production scaling.
Challenges we ran into
Getting the agents to sound genuinely human was harder than expected. Early outputs read like ads no matter how much we tuned the prompts. The breakthrough came from grounding generation in entity-level context from GLiNER rather than relying purely on LLM creativity, which kept responses specific and natural instead of generic and salesy. Orchestrating six independent agents with shared state across async pipelines introduced race conditions and data consistency issues that required careful coordination through the swarm manager. Integrating five external APIs (GLiNER via Fastino, Claude, Yutori, Reka, Senso) each with different authentication patterns, rate limits, and response formats meant building resilient service wrappers that degrade gracefully when any single dependency is unavailable. Building the real-time visualization of agent swarm activity required balancing WebSocket throughput with frontend rendering performance, especially when multiple campaigns run concurrent cycles.
Accomplishments that we're proud of
The self-evolution loop actually works. We can watch strategies that underperform get phased out while high-performing approaches get reinforced, all without human intervention. The GLiNER integration turned out to be the key architectural decision: by running zero-shot entity extraction at every pipeline stage, we achieved consistency that would have been impossible with prompt engineering alone. The knowledge graph in Neo4j gives us entity-level intelligence that goes beyond simple A/B testing. We can answer questions like which specific pain points drive the most engagement on which platforms, and feed that directly back into strategy generation. The Reka Vision integration adds a dimension most text-only marketing tools completely miss, giving agents the ability to understand visual context before engaging. The whole system runs as a genuine multi-agent swarm where each agent has a clear specialization but they all contribute to a shared, evolving understanding of what works.
What we learned
Zero-shot extraction is more powerful than fine-tuned models for this use case because the system needs to handle any product category without retraining. The knowledge graph approach to storing marketing intelligence is fundamentally better than flat analytics because relationships between entities, strategies, and outcomes capture patterns that aggregate metrics cannot. Human-sounding AI is not about better prompts; it is about better grounding. When agents have precise entity context from GLiNER and verified product knowledge from Senso, the outputs feel authentic because they are specific, not because they are trying to sound casual. Self-improving systems need fast feedback loops, and the combination of real-time metrics collection with Neo4j-backed pattern analysis creates a tight enough loop that the agents measurably improve across cycles.
What's next for Egg & Geese
Production deployment on Render with horizontal scaling so multiple campaign swarms can run independently. Expanding platform coverage beyond Twitter, Reddit, and Instagram to include TikTok comments, YouTube discussions, and niche forums. Building a marketplace where businesses can share and trade high-performing strategy templates across industries. Deeper Reka Vision integration for analyzing competitor visual content and generating image-aware engagement strategies. A full analytics dashboard with predictive modeling that forecasts which strategies will perform before they are deployed. Long-term, we want Egg & Geese to become the autonomous marketing layer that any business can plug into: describe your product, set your budget, and let the swarm handle the rest while you watch the knowledge graph grow smarter every day.
Log in or sign up for Devpost to join the conversation.