Agent Orchestration System

Production-grade reference implementation and tutorial for building agentic coding systems.

This project is intentionally runtime-agnostic. The same orchestration core can run on a local machine, Docker container, SSH host, browser-only worker, serverless runtime, or a full sandbox/VM.

What is inside

• Typed tool runtime with schema validation, permission checks, logging, and redaction.
• Agent loop that supports tool calls, event logs, cancellation, and provider adapters.
• Portable tools for filesystem, shell, Git, GitHub, web fetch, MCP, and LSP-style lookup.
• Subagent patterns for parallel repo reading and review-style workers.
• Runtime adapters that keep tools independent from the execution environment.
• Detailed docs explaining how to build your own system without using this code.

Package Map

packages/core          Agent loop, provider contract, events, subagent runner
packages/tool-runtime  Tool registry, schemas, permissions, runtime adapter API
packages/tools         Reusable production-style tools
packages/runtimes      Local/Docker/SSH/browser runtime adapter examples
packages/subagents     Repo-reader and review-worker orchestration patterns
examples               Small runnable integrations
docs                   Architecture and build-your-own tutorials

Quick Start

pnpm install
pnpm build

Read docs/00-overview.md first if you want the architecture. Read docs/13-building-your-own.md if you want to implement the concepts from scratch.

Design Goals

• Provider neutral: OpenAI, Anthropic, Gemini, local models, or any compatible adapter.
• Runtime neutral: no VM requirement.
• Tool-first: every capability is a typed, permissioned tool.
• Subagent-ready: bounded workers can run with restricted tools.
• Production-oriented: cancellation, retries, timeouts, event logs, redaction, and auditability.

"# agent-orchestration-system"