Agent Orchestrator
A lightweight orchestration layer for multi-agent LLM applications. Handles agent lifecycle management, inter-agent communication, and provides structured logging via OpenTelemetry.
Key features
- Declarative agent definitions using YAML configuration
- Built-in retry logic with exponential backoff for LLM API calls
- Shared memory via Redis for cross-agent state management
- Trace propagation across agent boundaries for end-to-end observability
Architecture
The orchestrator uses an event-driven architecture. Each agent runs as an independent coroutine, communicating through a central message bus. This design keeps agents loosely coupled while maintaining the ability to coordinate complex multi-step workflows.
Lessons learned
Building reliable multi-agent systems requires treating LLM calls as unreliable network calls. The same patterns that make distributed systems robust --- circuit breakers, bulkheads, timeouts --- apply directly to agent orchestration.