Excited to share an experimental project I've been working on: an agentic AI code assistant built entirely from scratch (no frameworks!).
Do you wanted to move beyond simple agent loops and explore the "Deep Agent" architecture style—the same advanced pattern powering tools like Claude Code, Manus and Deep Research.
This architecture excels at complex, multi-step tasks. My implementation features:
- A central Orchestrator Agent to analyse, plan, and delegate.
- A specialised Explorer Agent to investigate the codebase.
- A specialised Coder Agent to implement changes.
The key innovation? A shared Context Store.
This central "memory" allows agents to share findings, eliminating redundant work and creating a "compound intelligence" where the system gets smarter as it works. It's a fascinating model for strategic delegation and sustained reasoning.
Check out the Github project to learn more
Top comments (0)