I’m ready to move from concept to prototype on a self-organising multi-agent system that thinks, learns and acts with minimal hand-holding. At its core, the framework will spin up specialised agents that share context, retain long-term memory and coordinate through an internal reasoning loop I can inspect at any time. The first concrete agent I need online is an Optimizer that can ingest live Roadnet data, interpret routing constraints and iteratively improve plans. While the architecture should remain flexible enough to add planners or researchers later, the Optimizer must already demonstrate collaborative behaviour—asking for missing context, broadcasting findings and negotiating task ownership with its peers. Key capabilities I expect to see working end-to-end: • Context persistence across agent sessions so knowledge compounds over time • Transparent reasoning trace (I should be able to replay or audit what led to each decision) • Real-time monitoring through a lightweight local dashboard that visualises agent messages, memory lookups and task progress • Direct API integration with Roadnet for both read and write operations, handling authentication and data-format quirks gracefully You may choose the underlying stack—LangChain, LlamaIndex, custom Python micro-services, or another approach—so long as the final codebase is clean, documented and easy for me to extend. I’ll consider the project complete when I can: 1. Point the framework at a Roadnet instance, 2. Watch the Optimizer agent pull in routes, propose improvements and justify each step on the dashboard, and 3. Persist the entire interaction history for future runs. If you’ve built agentic systems before and enjoy turning abstract reasoning loops into reliable software, let’s talk.