Knowledge Broker

Inspiration The inspiration for Knowledge Broker came from a simple observation: On-chain data is abundant, but insight is scarce.

In the fast-paced world of the Base L2 ecosystem, users are drowning in dashboards, raw transaction logs, and endless social streams. We realized that the future of AI isn't just about chatting with a bot; it's about autonomous agents that can act as economic actors—scouts that go out into the digital wilderness, find the "alpha," and bring it back.

We wanted to answer a fundamental question: Can we build an AI agent that is smart enough to pay for itself?

This led us to the concept of the "Knowledge Broker"—an agent that doesn't just summarize news but acts as a Data Merchant. It uses Gemini 3 to synthesize high-value intelligence, teases it on social media (Farcaster), and uses crypto-native rails (x402 protocol) to transact that knowledge for USDC.

What We Learned Building with the Gemini 3 API family shifted our mental model of what an LLM can do. We moved from prompt engineering to "Reasoning Engineering."

• The Power of "Thinking": We learned that by exposing Gemini 3 Pro's internal reasoning process (the "Thought Stream"), we could trust the agent with complex, multi-step tasks. Watching the model "think" through a data discrepancy before posting gave us the confidence to let it run autonomously.
• Latency vs. Depth: We discovered that not every task needs a PhD. We learned to dynamically route queries: using Gemini 3 Flash for quick, high-energy social engagement and switching to Gemini 3 Pro when deep synthesis of conflicting data sources was required.
• The "Soul" of an Agent: We learned that an agent without a persona is just a script. By embedding a "Soul" file (SOUL.md) that defines the agent's principles, voice, and operational loop, we could maintain a consistent "Base Scout" identity across thousands of interactions.

How We Built It We architected the system as a dual-layer application: the Agent Core (Backend) and the Insight Interface (Frontend).

1. The Agent Core The brain of the operation is an event-driven loop powered by Gemini 3 Pro.

• Multimodal Input: The agent ingests structured SQL results from Dune Analytics and unstructured social sentiment from Farcaster.
• Tool Use: We built custom Python "Skills" (dune-data, farcaster-post) that the model invokes autonomously.
• Monetization: We integrated the x402 protocol to generate paywall links on the fly, allowing the agent to sell its generated JSON/CSV reports. Currently, the implementation is only on testnet so monetary values when testing.

1. The Insight Interface We built a real-time dashboard using Next.js, Tailwind CSS, and Framer Motion.
   • It subscribes to the agent's event stream to visualize the "Thinking Level" changes in real-time.
   • We used lucide-react to create a visual language for the agent's state (e.g., ⚡ for Flash, 🧠 for Pro reasoning).

Challenges We Faced

• The "Black Box" Problem: Initially, it was terrifying to let an AI post to a public social network autonomously. We didn't know why it made certain decisions. We solved this by building the Thought Stream visualization, which parses the raw logs to show the model's internal monologue before it takes action.
• Hallucination Risks: In crypto, a wrong number is dangerous. We had to implement a strict "Verification Step" in the agent's prompts, forcing Gemini 3 to cross-reference Dune data with a second source (DefiLlama) before declaring a trend.
• Handling Rate Limits: We hit the API limits frequently during high-intensity testing (specifically 429 errors). We had to architect a robust backoff strategy and a queue system that prioritizes "Pro" reasoning only for paid/high-value tasks, defaulting to "Flash" for general chatter.

Built With

• base-l2-blockchain
• defillama-api
• dune-analytics-api
• fastapi
• framer-motion
• google-gemini-3-flash
• google-gemini-3-pro
• lucide-react
• next.js-15
• neynar-api-(v2)
• node.js
• openclaw-agent-framework
• python-3.12
• tailwind-css-4
• typescript
• vercel
• x402-protocol