🤖 AgentBazaar: Emergent Skill Marketplace

AgentBazaar is a local multi-agent simulation where AI agents autonomously trade skills. It simulates a gig economy where agents post tasks, bid for work, negotiate contracts, execute deliverables, and validate results—all without human intervention.

Built using the Agno Framework (formerly Phidata) for orchestration and Ollama for local LLM inference.

🏗 Architecture

The system follows a strict linear orchestration flow designed to mimic high-stakes contracting environments:

graph TD
    User((User)) -->|Prompt| Broker[Broker Agent]
    Broker -->|TaskSpec| Workers[Worker Market]
    subgraph Market Logic
        Workers -->|Bids| Negotiator[Negotiator Agent]
        Negotiator -->|Winning Bid| Contract[Contract Agent]
    end
    subgraph Execution
        Contract -->|Contract JSON| Escrow[Escrow Agent]
        Escrow -->|Lock Funds| Executor[Executor Agent]
        Executor -->|Result| Validator[Validator Agent]
    end
    Validator -->|Pass/Fail| Escrow
    Escrow -->|Release/Refund| User
    Validator -->|Score| Reputation[Reputation DB]
    
    style User fill:#f9f,stroke:#333,stroke-width:2px
    style Broker fill:#fff,stroke:#333
    style Workers fill:#ccf,stroke:#333
    style Escrow fill:#dfd,stroke:#333

Loading

🧠 Components & Tech Matrix

1. Agents (Agno Framework)

We utilize Agno's Agent class to define specialized personas.

• Broker: Structurers messy user prompts into strict JSON TaskSpec.
• Workers: 3 distinct personas (Fast/Cheap, Premium, Balanced) that generate competitive bids.
• Negotiator: Implements a scoring algorithm (price vs reputation vs confidence) and runs a multi-turn negotiation loop to drive down prices.
• Validator: Acts as a strict QA, checking output against contract acceptance criteria.

2. LLM Engine (Ollama)

The system is built to run 100% locally.

• Model: llama3.2:latest (Swappable in src/agents/*.py)
• Inference: Zero-latency local calls via standard Ollama API.

3. State & Persistence

• Escrow Ledger: Simulated financial locking mechanism (json based).
• Reputation DB: Tracks long-term agent performance (Success Rate, Avg Score).

🚀 Getting Started

Prerequisites

• Python 3.9+
• Ollama installed and running.

Installation

1. Clone the repository
2. Install Dependencies

   pip install -r requirements.txt

3. Pull the Model

   ollama pull llama3.2:latest

Running the Simulation

Launch the visual dashboard:

streamlit run ui/app.py

💡 How It Works

1. Task Ingestion: You post "Write a clear Python function for finding primes".
2. Bidding War:
   • Worker A (Cheap) bids $40, Low Confidence.
   • Worker B (Premium) bids $120, High Confidence.
3. Negotiation: The Negotiator scores bids. It might challenge Worker B to lower their price to $100.
4. Contracting: A formal JSON contract is generated with "Acceptance Tests" derived from your prompt.
5. Execution: The winning agent writes the code.
6. Validation: The Validator runs a rubric check. If it passes, Escrow releases funds and Reputation increases.

🛠 Building with Agno

This project demonstrates key Agno patterns:

• Structured Outputs: Using output_schema to force LLMs to reply in valid Pydantic models (crucial for system reliability).
• Role Playing: Assigning distinct instructions to agents to create "Personas" that behave differently.
• Orchestration: Managing state hand-offs between multiple agents in a Pythonic workflow.

🔮 Future Use Cases

• Automated DevOps: Agents bidding on ticket resolution.
• Content Factories: SEO agents, Writers, and Editors negotiating details.
• Resource Allocation: Autonomous trading bots negotiating API usage.

Output Screenshots