Skip to content

Knowledge base

Multi-Agent and Orchestration

Multi-agent orchestration is the design of the system that dispatches, isolates, verifies and merges the work of many coding agents. Designed well, parallel agents multiply output. Designed badly, they multiply rework.

Written by

Updated:

Software Factories 5

Our position. We run a factory mostly on workflow, with agents only at the stations that need judgment; knowing which a step is tells you where to put a gate and where to trust the model (Workflows and agents (opens in a new tab)). The limit on fan-out is how much the team can check well, so set it there and raise it only when checking improves (A swarm is not throughput (opens in a new tab)).

The leverage in agentic coding moves from prompting one agent to designing how many agents share work. Orchestrators, autonomous loops, agent teams, swarms and runners are the building blocks. With any of them you have to settle when to fan out, how to keep concurrent work from colliding, and who checks the result.

The checking role is the one most sources agree on: an agent cannot grade its own work. A verifier in fresh context, ideally reading the real-world signal rather than the agent’s claims, catches what the creator talks itself out of (Creator and verifier roles). The factory that turns this coordination into autonomous production sits one layer up, in software factories and harness engineering.

Start here

Resources

Orchestration & Coordination

  • AgentField (opens in a new tab) - Open-source control plane for running AI agents as production infrastructure, with multi-language SDKs, agent mesh communication, and human-in-the-loop workflows.
  • Air (JetBrains) (opens in a new tab) - JetBrains agentic development environment that runs Codex, Claude Agent, Gemini CLI, and Junie side by side as independent parallel task loops, with delegation and oversight.
  • Augment Intent (opens in a new tab) - Augment Code’s developer workspace that orchestrates multiple AI agents around living specifications, delegating to parallel specialist agents and keeping specs synced with code.
  • Claude Flow (opens in a new tab) - Agent orchestration for Claude with multi-agent swarms, RAG integration, and native Claude Code and Codex hooks.
  • Claude Swarm (opens in a new tab) - Ruby gems for AI agent systems; SwarmSDK runs single-process orchestration with persistent memory, semantic search, and node workflows, plus SwarmMemory and SwarmCLI.
  • ClipboardHealth core-utils (opens in a new tab) - Clipboard Health’s open-source TypeScript monorepo bundling groundcrew (agent orchestrator using isolated git worktrees), ai-rules, playwright-reporter-llm, and production utilities.
  • Conductor (opens in a new tab) - Runs parallel Codex and Claude Code agents in isolated git worktrees, with a shared view of what each is working on for review and merge.
  • Gas City (opens in a new tab) - Orchestration-builder SDK for declarative multi-agent setups via city.toml manifests, with runtime providers, Beads-backed work tracking, and controller/supervisor loops.
  • Gastown (opens in a new tab) - Multi-agent workspace manager that coordinates 20-30+ Claude Code instances with persistent work queues, convoy-based task tracking, and crash recovery.
  • MCP Agent Mail (opens in a new tab) - Mail-like coordination layer for coding agents with asynchronous messaging, file reservation leases, searchable message history, and human oversight for shared codebases.
  • dmux (opens in a new tab) - Tmux pane manager that runs Claude Code, Codex, and OpenCode in parallel using git worktrees, with AI-generated branch names and single-keypress management.
  • fabro (opens in a new tab) - MIT-licensed Rust orchestrator that defines AI agent workflows as deterministic graphs in Graphviz DOT, with human approval gates and stylesheets routing tasks across models.
  • groundcrew (opens in a new tab) - Clipboard Health’s CLI that dispatches a ticket backlog to local AI coding agents, each isolated in its own git worktree and tmux pane; MIT.
  • herdr (opens in a new tab) - Rust terminal multiplexer for AI coding agents, showing each agent’s terminal output and status across panes, with SSH detach and reattach; AGPL-3.0-or-later, commercial option.
  • kagent (opens in a new tab) - Kubernetes-native agent runtime from the Istio founders that defines agents as CRDs managed with GitOps, RBAC, and OpenTelemetry, working with any LLM or framework.

Autonomous Loops & Runners

  • Loop Engineering (opens in a new tab) - MIT-licensed reference repo of patterns, starters, and Node CLI tools for building autonomous loops that orchestrate AI coding agents until a goal completes.
  • Millwright (opens in a new tab) - MIT-licensed autonomous agent orchestrator in Ruby that polls a GitHub Project board and dispatches Claude Code to implement issues and open PRs.
  • Open Agents (opens in a new tab) - Vercel-labs open-source tool that spawns coding agents in isolated cloud sandboxes with their own git branches, explorer and executor subagents, and durable resumable loops.
  • Ralph (opens in a new tab) - Autonomous engineering loop that runs Claude Code as relay sessions; each session completes one task, commits, and triggers the next via bash.
  • Ralph loop (opens in a new tab) - Autonomous loop running Amp or Claude Code until all PRD items complete; each iteration starts with clean context and persists state via git.
  • Ralphy (opens in a new tab) - Bash script for autonomous development that reads tasks from PRDs, YAML, or GitHub Issues and dispatches them to Claude Code, OpenCode, Codex, or Cursor.

Multi-Agent Frameworks

  • Agent Workflow Protocol (opens in a new tab) - Open standard for multi-agent workflows from scripted pipelines to self-organizing teams; declarative, runtime-agnostic, portable.
  • AgentFlow (Stanford) (opens in a new tab) - Trainable tool-integrated agentic framework with four modules coordinated through evolving memory, using Flow-GRPO to optimize the planner so a 7B backbone competes with larger models.
  • Aperant (opens in a new tab) - Multi-agent coding framework (formerly Auto-Claude) with a desktop app running up to 12 parallel agent terminals, git worktree isolation, and self-validating QA loops.
  • Buford (opens in a new tab) - Spec-driven AI development template for Claude Code with multi-agent and human-in-the-loop workflows.
  • CrewAI (opens in a new tab) - Framework for orchestrating role-playing autonomous agents into collaborative crews with tasks and tools.
  • CrewAI Examples (opens in a new tab) - Collection of example projects showing how to build multi-agent crews and sequential or hierarchical flows with the CrewAI framework to automate workflows.
  • DeerFlow (opens in a new tab) - ByteDance’s open-source long-horizon agent harness combining LangGraph orchestration with sandboxes, memory, tools, skills, subagents, and a message gateway across web search, crawling, and code execution.
  • Equilateral Agents (opens in a new tab) - MIT-licensed framework with 22 self-learning agents, a three-tier standards system, and a Knowledge Synthesis Flywheel that tracks execution history and extracts patterns from failures.
  • Hermes Agent (opens in a new tab) - Self-improving agent framework by Nous Research that creates skills from experience and refines them in use, with persistent memory and Telegram and Discord bridges.
  • LangChain (opens in a new tab) - Framework for building LLM applications and agents by composing models, tools, memory, and prompts.
  • LangGraph (opens in a new tab) - LangChain’s library for building stateful, multi-actor agent workflows as graphs.
  • LatentMAS (opens in a new tab) - Reference implementation of latent-space multi-agent collaboration where agents share internal representations through a latent working memory instead of exchanging text.
  • Legacy Modernization Agents (opens in a new tab) - Microsoft framework that converts legacy COBOL to Java Quarkus or C# .NET, with reverse-engineered business logic extraction, Neo4j dependency mapping, and a review portal.
  • Microsoft Autogen (opens in a new tab) - Microsoft’s programming framework for agentic AI with a layered architecture, multi-language support, and AutoGen Studio GUI.
  • PocketFlow (opens in a new tab) - 100-line LLM framework for building agents, workflows, and AI applications without vendor lock-in.
  • Pydantic AI (opens in a new tab) - Agent framework from the Pydantic team for building type-safe LLM apps with structured, validated outputs.
  • Semantic Kernel (opens in a new tab) - Microsoft’s open-source SDK for integrating LLMs into apps with plugins, planning, and memory.
  • SWE-AF (opens in a new tab) - Multi-agent software engineering system on AgentField that coordinates planning, execution, and governance agents across models, with parallel execution via isolated git worktrees.
  • smolagents (opens in a new tab) - Hugging Face’s minimal library for building agents that act by writing and executing code.

Agent Collections & Libraries

Agent Design Patterns

In this section

  • Creator and verifier roles

    Patterns that separate making from checking, why self-grading fails, and how to wire a verifier into the loop, for teams running more than one agent.