A highly scalable, multi-agent artificial intelligence research and engineering assistant. Lyra orchestrates four specialized agents (Researcher, Engineer, Planner, and Critic) to execute complex, multi-step tasks.
Built for enterprise-scale from day one, Lyra features an asynchronous event-driven architecture. It completely decouples agent logic from network connections. This design allows independent scaling and fault tolerance while using a federated model strategy to maximize throughput on free-tier interfaces.
To avoid rate limits while maintaining high performance, Lyra assigns specific agents to dedicated application programming interfaces.
- Planner:
gpt-oss-120bvia Cerebras. Breaks down tasks with high reasoning capability at fast speeds without exhausting primary rate limits. - Researcher:
DeepSeek V4 Flashvia OpenRouter. Uses its massive context window to process large documents and run complex data retrieval pipelines. - Engineer:
MiniMax M2.5via OpenRouter. Excels at autonomous coding and tool usage within the execution sandbox. - Critic:
Llama 3.3 70Bvia Groq. Delivers rapid reviews, ensuring the final quality check does not delay the final output.
Lyra separates networking infrastructure from cognitive agent execution using an event-driven message bus.
- Technology: Go with the Gin framework.
- Role: Manages authentication, WebSockets, and file uploads. It pushes events to the message broker and handles thousands of persistent connections with minimal memory usage.
- Technology: Apache Kafka.
- Role: Decouples agents. The Planner publishes a task, which the Researcher consumes at its own pace. This ensures reliable message delivery and robust retry logic.
- Technology: Python, FastAPI, and LangGraph.
- Role: Each agent runs as an isolated, containerized microservice. This allows independent scaling for resource-heavy agents like the Researcher.
- Technology: Go, Docker, and gVisor.
- Role: Provides a strict, kernel-isolated environment for the Engineer agent to safely compile and test generated code.
- Relational and Graph State: PostgreSQL. Handles agent state checkpointing and user account data.
- Vector Search: pgvector. A native PostgreSQL extension for storing and querying text embeddings.
- Artifact Storage: AWS S3. Stores user documents and generated code artifacts.
Ingest: The user submits a prompt via the frontend interface.
Route: The Go API Gateway validates the security token and publishes a task event to Apache Kafka.
Plan: The Planner Agent consumes the event, creates a step-by-step plan, and publishes a research request.
Execute: The Researcher and Engineer agents consume their respective tasks. Engineer code is routed to the Go Sandbox worker.
Review: The Critic Agent consumes the execution outputs and verifies accuracy.
Stream: The Go Gateway consumes the final completion event and streams the text to the user via WebSockets.
License
MIT