𝗪𝗵𝗮𝘁 𝗜 𝗟𝗲𝗮𝗿𝗻𝗲𝗱 𝗕𝘂𝗶𝗹𝗱𝗶𝗻𝗴 𝗮𝗻 𝗔𝘂𝘁𝗼𝗻𝗼𝗺𝗼𝘂𝘀 𝗗𝗲𝗮𝗹 𝗛𝘂𝗻𝘁𝗶𝗻𝗴 𝗔𝗴𝗲𝗻𝘁 𝗦𝘆𝘀𝘁𝗲𝗺

𝗪𝗵𝗮𝘁 𝗜 𝗟𝗲𝗮𝗿𝗻𝗲𝗱 𝗕𝘂𝗶𝗹𝗱𝗶𝗻𝗴 𝗮𝗻 𝗔𝘂𝘁𝗼𝗻𝗼𝗺𝗼𝘂𝘀 𝗗𝗲𝗮𝗹-𝗛𝘂𝗻𝘁𝗶𝗻𝗴 𝗔𝗴𝗲𝗻𝘁 𝗦𝘆𝘀𝘁𝗲𝗺

I built a multi-agent AI system to find product deals.

It scans the internet for bargains. It estimates the true value of products using three different methods. It sends a notification to my phone when it finds a deal worth buying.

Here is what I learned about building agentic systems.

𝟭. 𝗦𝗽𝗲𝗰𝗶𝗮𝗹𝗶𝘇𝗲 𝗮𝗻𝗱 𝗢𝗿𝗰𝗵𝗲𝘀𝘁𝗿𝗮𝘁𝗲 No single model is good at everything. I split the work into small, focused agents:

• Scanner Agent: Finds deals from RSS feeds.
• Specialist Agent: A fine-tuned Llama 3.2 model for price estimation.
• Frontier Agent: A GPT model using RAG for context.
• Ensemble Agent: Blends different estimates into one final number.
• Planning Agent: The brain that decides which tool to use and when.

The intelligence comes from how they work together, not from one giant prompt.

𝟮. 𝗥𝗔𝗚 𝗶𝘀 𝗖𝗵𝗲𝗮𝗽 𝗟𝗲𝘃𝗲𝗿𝗮𝗴𝗲 Adding a few similar products and their prices to a prompt improved accuracy. I used a Chroma vector database to find these products. Seeing product categories cluster in a 3D map proved that embeddings work.

𝟯. 𝗘𝗻𝘀𝗲𝗺𝗯𝗹𝗶𝗻𝗴 𝗪𝗼𝗿𝗸𝘀 𝗕𝗲𝘁𝘁𝗲𝗿 I did not pick one favorite pricing method. I used a weighted blend:

• 80% weight to the RAG model.
• 10% weight to the fine-tuned model.
• 10% weight to a neural network. This blend outperformed any single model.

𝟰. 𝗗𝗲𝗳𝗲𝗻𝘀𝗶𝘃𝗲 𝗧𝗼𝗼𝗹 𝗖𝗮𝗹𝗹𝗶𝗻𝗴 LLMs often send malformed data when calling functions. Smaller models might add extra arguments that do not exist. You must write code that filters these arguments against your actual function signature. This prevents your system from crashing at night.

𝟱. 𝗠𝗮𝗻𝗮𝗴𝗲 𝗜𝗻𝗳𝗿𝗮𝘀𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗲 𝗦𝗲𝗿𝘃𝗲𝗿𝗹𝗲𝘀𝘀𝗹𝘆 I used Modal to run my fine-tuned models. You can request a GPU with one line of code. This removes the need to manage your own servers. You can also deploy a model as a remote service that you call like a normal Python object.

𝟲. 𝗠𝗮𝗸𝗲 𝘁𝗵𝗲 𝗨𝗜 𝗙𝗲𝗲𝗹 𝗔𝗹𝗶𝘃𝗲 An agent can feel like a black box. I used Gradio with background threads and generators. This allows the UI to show live logs while the agent works. When you see the agent thinking in real time, you trust it more.

Building this taught me that agentic magic is mostly just good plumbing. Get your data flow right, use structured outputs, and keep your agents focused.

Source: https://dev.to/m_toqeer/what-i-learned-building-an-autonomous-deal-hunting-agent-system-3n6b

Optional learning community: https://t.me/GyaanSetuAi