Whatsapp Agent
A capable, Whatsapp-based AI Agent
Overview
A Whatsapp-based AI Agent that expands on Claude's basic functionality using tools like image gen, editing and web search. Features adaptive context, tailoring the Agent to my needs over time by editing the system prompt through a tool call. Designed in a layer-by-layer system, so that it can be expanded on and improved over time.
Tech Stack
Python
FastAPI
PostgreSQL
Claude
Key Decisions
1.
Layered Architecture
In order to improve the model's capabilities over time, I had to properly plan an architecture. This architecture is layered, e.g. Layer 1 - Engine (which controls the decision making and runs the Agent). If I plan to make more Agents, I can then use this same architecture for them.
2.
Self-learning Context
Had to figure out a way to have the Agent edit sections of its own system prompt in order to store new knowledge for future use. Settled on just simply making a tool for it, as I know that LLMs are great at deciding when to use tools.
3.
Tool Selection
The next step to expanding on existing LLM models is to add tools that they do not have access to, turning it into an LLM harness. I use image generation tools often, so I implemented OpenAI's image tool along with editing functionality. Plan to add many more tools over time.