12Agent Harnesses: Hermes, Claude Code & More
Stop writing agent loops from scratch. Install Hermes Agent, connect it to a local Ollama model, give it real work and reach it from Telegram — plus how to choose among other harnesses.
What you'll learn
- Explain what an agent harness provides
- Install and configure Hermes Agent with a local or cloud model
- Use skills, memory, scheduling and the messaging gateway
- Compare Claude Code, OpenHands, Aider, CrewAI and LangGraph
What is a harness?
In Lesson 09 you wrote an agent loop by hand. A harness is production-ready software that provides everything around the model: the loop, a toolbox (files, terminal, web, browser), context management, memory, permissions, scheduling and interfaces (terminal, desktop, chat apps). You bring the model and the goals.
| You write yourself | A harness gives you |
|---|---|
| The think → act → observe loop | A tested loop with retries, context compaction and limits |
| Every tool | Built-in file, shell, web search, browser and MCP support |
| Memory | Persistent memory and reusable skills |
| An interface | CLI/TUI, desktop app, and messaging gateways |
| Scheduling | Built-in scheduled tasks and delivery |
Hermes Agent (Nous Research)
Hermes Agent is an open-source (MIT-licensed) personal agent. Its standout feature is a learning loop: it creates skills from experience, improves them as it uses them, keeps persistent memory across sessions, and can run scheduled tasks and reach you on Telegram, Discord, Slack, WhatsApp, Signal or email. It works with cloud providers or any OpenAI-compatible endpoint — including your own Ollama.
Step 1 — Install
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
source ~/.bashrc # or: source ~/.zshrciex (irm https://hermes-agent.nousresearch.com/install.ps1)Prefer an app? Download Hermes Desktop from the Hermes website; it bundles the agent and its runtime.
The installer sets up its own Python, Node.js and other dependencies. Run hermes doctor any time to diagnose problems.
Step 2 — Choose a model
Hermes needs a model with a large context window — its docs call for at least 64,000 tokens. Ollama's default context is much smaller, so raise it before connecting:
# pick a strong tool-calling model that fits your hardware
ollama pull qwen3:14b
# start Ollama with a 64K context window
OLLAMA_CONTEXT_LENGTH=65536 ollama serveThen, in another terminal:
hermes model
# → choose "Custom Endpoint"
# → Base URL: http://localhost:11434/v1
# → Model: qwen3:14bhermes model # pick a provider (OpenRouter, OpenAI, and others) and paste your API key
# or the one-step option from Nous Research:
hermes setup --portal # OAuth login + model + web search, image, TTS & browser toolsStep 3 — Start working
hermes # classic CLI
hermes --tui # modern terminal UI
hermes -c # continue your last sessionTry real tasks:
› Read ~/clients/kape/discovery-notes.md and draft a one-page website proposal
in proposal.md using my usual structure: goals, scope, timeline, price.
› Search the web for 5 cafés in Iloilo with no online ordering, and save
name, link and notes to ~/leads/cafes.csv
› /tools (see what tools are enabled)
› /model (switch models mid-session)Step 4 — Skills & memory
When Hermes completes a multi-step task, it can save the procedure as a skill — a reusable instruction file it reloads next time a similar task appears, and refines as it goes. Its memory stores facts about you and your projects across sessions. Skills follow the open Agent Skills format, so you can also write or share them.
Step 5 — Reach it from your phone
hermes gateway setup # connect Telegram, Discord, Slack, WhatsApp, Signal, email…With the gateway running (on your computer or a small always-on server), you can message your agent: “Every Monday 8am, summarize last week's leads and send them here.” Hermes supports scheduled tasks with delivery to your chat platform. See the messaging guide for platform-specific steps.
Step 6 — Add tools with MCP
Hermes can connect to MCP servers — like the invoice server you built in Lesson 11. Follow the Hermes MCP guide to register servers, then control which tools each platform may use with hermes tools.
Other harnesses worth knowing
Claude Code
Anthropic's agentic coding tool for the terminal and IDE: reads your codebase, edits files, runs commands, uses MCP and skills. The same harness is available as the Claude Agent SDK for building your own agents.
OpenHands
Open-source software-engineering agent platform that works in a sandbox: writes code, runs tests, browses documentation. Docs.
Aider
Terminal pair-programmer that edits files in your Git repo and commits each change. Works well with local models: aider --model ollama_chat/qwen3:14b. Docs.
CrewAI
Python framework for role-based agent teams (researcher, writer, analyst) that collaborate. Docs.
LangGraph
Low-level framework for stateful agents as graphs, with checkpoints and human-in-the-loop. Choose it for precise control. Docs.
OpenAI Agents SDK / Pydantic AI
Lightweight code-first SDKs for building agents with tools, handoffs and guardrails. OpenAI Agents SDK · Pydantic AI.
Which should you use?
| If you want… | Start with |
|---|---|
| A personal, always-on assistant that learns and messages you | Hermes Agent |
| To build and fix websites or code faster | Claude Code, Aider, OpenHands |
| To ship an agent inside your own product | Claude Agent SDK, OpenAI Agents SDK, Pydantic AI |
| Precise control over a complex business process | LangGraph |
| A team of role-based agents | CrewAI |
| No code at all | n8n AI Agent node, Dify, Flowise (Track 3) |
Key takeaways
- A harness provides the loop, tools, memory, scheduling and interfaces so you only bring the model and goals.
- Hermes Agent: open source, self-improving skills, persistent memory, scheduled tasks and chat gateways.
- Local models for Hermes need a big context window (64K+) and serious hardware; cloud models are the easier start.
- Pick harnesses by job: Hermes for personal ops, Claude Code/Aider/OpenHands for code, SDKs/LangGraph for products.
Knowledge check
0 / 3Q1What does an agent harness provide?
Harnesses package everything except the model weights and your goals.
Q2Minimum context window Hermes Agent's docs call for?
Hermes requires at least 64K context; raise Ollama's context length accordingly.
Q3Which command connects Hermes to Telegram and other chat apps?
The gateway setup connects messaging platforms.