Back

How-to5 min read

How to choose an agentic harness

Guide card: How to choose an agentic harness

You want an AI agent that does real work on your machine or answers you from a chat app, and you keep hearing the names Hermes, OpenClaw and Pi. Pick the harness by where you want to talk to the agent and what it must remember: OpenClaw for one assistant reachable from many chat apps, Hermes Agent for an assistant that keeps learning, Pi for a small coding harness you script and extend. All three are open source, install with one command, and work with many model providers.

What a harness is

The model only turns text into text. The harness is the runtime around it: the loop that sends a prompt, reads the reply, calls a tool, feeds the result back and repeats until the task is done. It also decides which tools exist, what is saved between sessions, where messages come in from, and what the agent is allowed to touch. Swap the model and the harness stays. Swap the harness and the same model behaves like a different product.

Step by step

  1. Decide where you talk to the agent. If you want to message it from your phone in Telegram, Slack or WhatsApp, you need a harness with a gateway: OpenClaw or Hermes. If the agent works in a terminal, a script or inside your own app, Pi is the lighter fit.

  2. Decide what it must remember. Hermes Agent from Nous Research is built around a learning loop: agent-curated memory, skills it writes after complex tasks and improves during use, and search over past conversations. OpenClaw is a self-hosted Gateway that holds sessions, routing and channel connections, with skills and plugins on top. Pi (PI Agent) keeps sessions as branchable trees and leaves long-term memory, sub-agents and plan mode to extensions.

  3. Set up Hermes Agent. On Linux, macOS or WSL2 (installation guide):

    Hermes Agent mark redrawn in Rebels engraved style: a portrait of a girl

    curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
    source ~/.bashrc    # or: source ~/.zshrc
    hermes model        # choose a provider and model
    hermes              # start chatting in the terminal

    To reach it from a chat app, run hermes gateway setup and then hermes gateway start. On macOS and Windows there is also a desktop app, and native Windows has a PowerShell installer.

  4. Set up OpenClaw. On macOS, Linux or WSL2 (getting started):

    OpenClaw mark redrawn in Rebels engraved style: a lobster mascot next to the OpenClaw wordmark

    curl -fsSL https://openclaw.ai/install.sh | bash
    openclaw gateway status
    openclaw dashboard

    The installer sets up Node.js if needed and starts onboarding. Choose Quick start to reuse an existing provider login or API key and keep the Gateway running in the background. The dashboard opens the Control UI in your browser; send a message there first, then connect a channel. The docs name Telegram as the fastest, since it only needs a bot token.

  5. Set up Pi. On macOS or Linux (quickstart):

    Pi mark redrawn in Rebels engraved style: a letter P built from hatched blocks

    curl -fsSL https://pi.dev/install.sh | sh
    pi --version
    cd /path/to/project
    pi

    Inside Pi, run /login to connect a subscription or API key, and /model to switch models. For scripts, pi -p "query" runs one task and exits; pi --continue resumes the last session in that folder. RPC and a TypeScript SDK cover embedding it in another program.

  6. Draw the safety boundary before you connect anything. OpenClaw runs tools on the host for the main session unless you configure sandboxing, and pairs unknown direct-message senders by default. Pi has no built-in permission system and runs with your user’s permissions; its docs recommend a container or micro-VM for untrusted or unattended work. Hermes has command approval, DM pairing and container backends, covered in its security guide. Every message, web page and file the agent reads may carry instructions, so give it the fewest tools it needs.

How the three relate

They are not three unrelated products. Pi names OpenClaw as a real-world integration of its SDK, and the OpenClaw README thanks Pi. OpenClaw treats models and harnesses such as Claude, Codex or local models as swappable plugins. Hermes can import settings, memories and skills from an OpenClaw install; preview it first with hermes import-agent openclaw --dry-run after installing the agent-importers plugin. Skills written as SKILL.md folders move between them more easily than anything else you build.

Common mistakes

  • Choosing by the model instead of the harness. The same model in a gateway assistant and in a terminal harness gives you very different tools, memory and risk.
  • Connecting a chat channel on day one. Get the agent working in the terminal or dashboard first, then open it to messages.
  • Running unattended on your own user account. Put scheduled or chat-driven work in a container or sandbox backend.
  • Treating memory as free. What the agent saves shapes every later answer. Read and prune it like any other config.
  • Installing everything at once. Start with one harness and one task. Move only when you hit a limit you can name.

Next step

Write down one task you want an agent to do, where the request will come from, and what it must remember between runs. Those three answers point at one of the harnesses above. If you are still choosing that task, see how to choose the first process for an AI agent.

If you want to talk through an agent setup for your own team, book a call on this page to start with a Discovery workshop, or begin from the Rebels Software homepage.

Sources

Logos belong to their respective projects and are shown only to identify them.

  • AI agents
  • Agentic harness
  • Hermes Agent
  • OpenClaw
  • Pi