MAmatiasmolinas·Mar 29, 2025
Hey HN, quick update on the Evolving Agents toolkit (https://github.com/matiasmolinas/evolving-agents). Thanks again for all the great feedback on the original thread!
Following up on the feedback and refactoring mentioned previously, I've focused on solidifying the architecture based on the agent-centric philosophy. The key components and their interactions are now detailed here: https://github.com/matiasmolinas/evolving-agents/blob/main/d...
In short:
SystemAgent (ReAct): Orchestrates using specialized tools, not hardcoded logic.
ArchitectZero: Designs solutions, outputting a structured JSON design.
SmartLibrary: Manages component persistence and semantic search (using ChromaDB).
SmartAgentBus: Handles capability-based discovery and routing between agents/tools (basic implementation working).
Workflow Lifecycle: Clear Design -> Generate
Read the rest
-> Process -> Execute flow managed by the SystemAgent.
The recent discussions on ACP (https://github.com/orgs/i-am-bee/discussions/284) have been super insightful and are guiding the next steps:
Dual Bus Architecture: Based on @ismaelfaro's suggestion, I'm planning to split the SmartAgentBus implementation into a dedicated system bus (for metadata, discovery, system-level messages) and a data bus (for agent-to-agent payload communication). This feels like a cleaner way to manage the different communication needs.
Smart Context/Memory: Addressing concerns raised by @QuantGeekDev and others about context pollution and determinism, I'll be exploring concepts for smart memory and smart context. The goal is to develop mechanisms that help agents pass only the most relevant context during interactions over the agent bus, potentially using prioritization or summarization.
I also evaluated how MCP might fit in. Components like the SmartLibrary or SmartAgentBus could conceptually act as MCP servers, while agents like SystemAgent or ArchitectZero could be MCP clients. However, I ran into a fundamental question: If agents themselves can also act as tools (like in the OpenAI Agents SDK), does requiring potentially numerous MCP servers (one per agent/tool instance) scale well? Or is a different abstraction needed for direct agent/tool comms?
Given this, my immediate focus will be on refining the core communication patterns (dual bus, smart context) before deciding on the best way to integrate MCP specifically for agent-to-agent communication.
This is all evolving (pun intended!), and I'm actively working on the dual bus and smart context POCs now. You can see the current architecture details in the ARCHITECTURE.md link above. Keen to hear your thoughts on this direction!