Orchestrator_and_Harness

Agent Harness

• Agent harnesses are opinionated, batteries-included frameworks with built-in tools and capabilities for building sophisticated, long-running agents
• Agent Harness SDK
  • Deep Agents SDK (Langchain)
  • Claude Agent SDK
  • CrewAI
  • AutoGen by Microsoft (maintenance mode)
  • Strands Agent by AWS (only AWS, less focus today)
• Agent Harness Applications
  • OpenClaw
  • Hermes
  • AutoGPT (less focus today)
  • AgentGPT (less focus today)
  • Claude Code
  • Codex
  • GitHub Copilot
• https://huggingface.co/learn/context-course/unit6/introduction
• https://docs.langchain.com/oss/python/concepts/products#agent-harnesses-like-the-deep-agents-sdk
• https://docs.langchain.com/oss/python/deepagents/overview
• https://www.youtube.com/watch?v=hcm5zIWASCM (Strands)
• https://medium.com/@a.posoldova/comparing-4-agentic-frameworks-langgraph-crewai-autogen-and-strands-agents-b2d482691311

Agent Harness Features

• Planning
• Context Management
• State Management
• Memory
  • Short Term, Long Term
• Skills, Tools and MCP
• Shell Execution
• Interpreter
• Sub-Agents
• Human-in-the-loop (HITL)
• Multi-step workflow
• Long running tasks
• Error Recovery
• Time Travel (Prompt Editing + Replay)

Agent Orchestrator

• https://www.ibm.com/think/topics/ai-agent-orchestration
• https://github.com/resources/articles/what-is-ai-agent-orchestration
• It is the process of coordinating multiple specialized sub-agents to achieve given objective
• In a way this also has same meaning as Agent Harness

OpenClaw and Hermes

• https://trilogyai.substack.com/p/technical-deep-dive-hermes-vs-openclaw
• https://github.com/NousResearch/hermes-agent
• https://openclaw.ai/
• Philosophy
  • Focus on routing and control
  • Focus on memory and self improvement
• Architecture
  • OpenClaw: Node.js
    • Good for I/O heavy gateway work
  • Hermes: Python
    • Access to ML ecosystem, Self Learning loop
• Session History
  • Both uses SQLite with FTS5
  • Hermes has bounded Memory and audits its own memory
  • OpenClaw stores memory in markdown files and can be edited
• Execution Environment
  • OpenClaw runs the heavy work on local
  • Hermes can offload heavy work to serverless execution like backends like Daytona/Modal
• Both have multiple chat gateways including WhatsApp, Discord, Telegram
• Both have Skills support and come with preinstalled skills

LangGraph

• https://docs.langchain.com/oss/python/langgraph/overview
• LangGraph provides low-level supporting infrastructure for any long-running, stateful workflow or agent
• Agent is composed of steps called Nodes which connect to form graph
  • START node for start
  • END node for end
• All nodes have a shared state
• Types of steps/nodes
  • LLM step
    • analyze, generate text or decisioning
  • Data step
    • Retrieve information
  • Action step
    • Perform external action
  • User input step
    • Need human intervention
• Nodes can include error handler and retry policy
• Common design
  • https://docs.langchain.com/oss/python/langgraph/workflows-agents
  • Prompt Chaining
  • Parallelization
  • Orchestrator worker
  • Agents
• Langchain Agents are internally implemented using LangGraph