Overview
Single-agent loops degrade on multi-part questions: the context fills with intermediate work, and the model loses track of what it was originally asked.
This system borrows the Plan → Reason → Act structure from Vision-Language-Action architectures and applies it to research. Decomposing the question and executing the parts become separate concerns, handled by separate agents.
Architecture
- Orchestrator: decomposes the incoming query and routes each part to a specialist instead of answering it directly.
- Research and Reasoning specialists: handle retrieval-shaped and inference-shaped subtasks, then hand results back for synthesis.
- Dual-layer memory: ChromaDB provides semantic long-term recall, and a sliding working-memory window with pruning keeps the active context bounded.
- Tool registry: Calculator, ArXiv and Search exposed through OpenAI-compatible function calling, so each tool is declared once and callable by any agent.
- Fully async: built on asyncio, so agent operations run concurrently instead of queuing behind each other.
Interface: explicit modes over inferred intent
It runs as an interactive CLI with explicit verbs: search for research, reason for analysis, and stats to inspect memory and system state.
Making the agent’s mode a command rather than something the model infers keeps the system debuggable: when a result is wrong, you know which path produced it.