I built an open-source orchestration engine for multi-agent AI – Fulcrum

Been thinking about the same problem everyone building multi-agent systems runs into: you spend more time writing routing logic, retry handlers, and state management than actually building agents.

So I built Fulcrum — a TypeScript library that handles the plumbing:

• Task decomposition via LLM (splits your task into parallel subtasks automatically)
• Agent routing by capability type (researcher, writer, analyst, coder, etc.)
• Parallel execution with dependency graphs (tasks that don't depend on each other run concurrently)
• Automatic retry with exponential backoff
• SQLite persistence so state survives process restarts

Works with Anthropic and OpenAI out of the box:

import { Fulcrum, createAnthropicAgents, createAnthropicDecomposer } from '@fulcrumhq/core'

const orchestrator = new Fulcrum({
  agents: createAnthropicAgents({ apiKey: process.env.ANTHROPIC_API_KEY }),
  decompose: createAnthropicDecomposer({ apiKey: process.env.ANTHROPIC_API_KEY }),
})

const result = await orchestrator.run(
  'Research our top 3 competitors and draft a positioning doc'
)

console.log(result.summary)

npm install @fulcrumhq/core

GitHub: https://github.com/hashegsethbeenfired-web/fulcrum

Would genuinely love feedback from anyone building multi-agent pipelines. What's the biggest pain point you hit that this doesn't solve?