The Problem
Building AI agents is straightforward. Building systems of agents that reliably work together is not. You need:
- Structured communication between agents
- Type-safe tool definitions
- Delegation patterns that don't lose context
- Fallback strategies when an agent fails
The Approach
Prometheus provides primitives for multi-agent orchestration:
Agent Definition
Define agents with typed capabilities, tools, and system prompts. Each agent knows what it can do and what it should delegate.
Workflow Composition
Compose agents into workflows with explicit data flow. Outputs from one agent feed into the next, validated by Zod schemas at every boundary.
Delegation
Agents can spawn sub-agents for specialized tasks. The parent agent maintains context and synthesizes results. Failed sub-agents don't crash the workflow — they return structured errors.
Design Philosophy
- Type safety everywhere — Zod schemas at every agent boundary
- Fail gracefully — structured errors, not crashes
- Composable — small agents that do one thing well
- Observable — every agent action is logged and traceable