AI agent harnesses compared: Claude Code, Codex, OpenCode, Pi, Hermes, OpenClaw and T3 Code

“Which AI agent should we use?” is one of the questions we hear most. The honest answer is that it depends on the job, because the tools people compare aren’t really competing. Some write code in a repository. Some run all day on a server and answer you on Telegram. One doesn’t do any work itself and just manages the others.

All of them are harnesses: the software around a model that gives it tools, context, permissions and a loop. (If that term is new, start with What is an agent harness?) This guide sorts the main ones into three groups and explains what each is good at.

These projects change monthly. Everything here is accurate at the time of writing; check each project’s own documentation before you commit.

The three kinds of harness

Kind	Examples	Lives where	You talk to it through
Coding agents	Claude Code, Codex, OpenCode, Pi	In a repository, in a terminal or editor	A terminal, IDE or app
Self-hosted personal agents	Hermes Agent, OpenClaw	On your machine or server, always on	Chat apps: Telegram, Slack, WhatsApp…
Control surfaces	T3 Code	On a machine that runs the coding agents	Desktop, web and mobile apps

Coding agents

Claude Code

Anthropic’s coding agent. It runs in the terminal, in VS Code and JetBrains, in a desktop app and on the web, reads and edits your code, runs commands, and asks before risky actions.

Why it’s good: Tight integration with Claude models, and the most mature extension system of any coding agent: CLAUDE.md for project context, skills for on-demand procedures, hooks for rules that must always run, subagents for side tasks, and MCP for outside systems. That makes it easy to encode a team’s way of working. See Claude Code skills, hooks and subagents.

Watch out for: It’s built around Claude models, and heavy use needs a paid plan or API spend.

Choose it if: you want a polished agent you can shape to your team’s conventions.

Codex

OpenAI’s coding agent, available as a CLI, an app and in the cloud, signed in with your ChatGPT account (codex login) or an API key.

Why it’s good: Strong with OpenAI’s coding models, available in several forms (local CLI, desktop app, cloud tasks), and widely supported by other tools, including T3 Code.

Watch out for: It is built around OpenAI models. Its extension model differs from Claude Code’s, so team conventions don’t transfer one-to-one.

Choose it if: your team already lives in the OpenAI ecosystem, or you want a second strong agent alongside Claude Code.

OpenCode

An open-source (MIT) coding agent with a terminal interface, plus a desktop app in beta.

Why it’s good: It works with many model providers and local servers, so you aren’t tied to one lab, and you can move a task to a cheaper or local model. It ships with a build agent for full-access development, a read-only plan agent for exploring code safely (switch with Tab), and a general subagent for complex searches. One independent study of coding harnesses ranked it highly for efficiency.

Watch out for: More choice means more configuration; results depend heavily on which model you pair it with.

Choose it if: you want open source, provider freedom or local models.

Pi

A deliberately minimal coding agent. Its README calls it “a minimal, extensible agent harness that you can make your own.”

Why it’s good: It ships with strong defaults and leaves out features like subagents and plan mode, so you add only what you need, through extensions, skills, prompt templates and shareable packages. It can also be scripted (print or JSON mode, RPC, a TypeScript SDK) and works with several model providers.

Watch out for: Minimal means you build more yourself.

Choose it if: you like understanding and owning every part of your tools.

Self-hosted personal agents

Hermes Agent

An open-source (MIT) agent from Nous Research that runs as a long-lived process on your machine or a server and talks to you through Telegram, Discord, Slack, WhatsApp, Signal and more.

Why it’s good: It takes continuity seriously. It keeps memory across sessions, builds a model of you over time, turns repeated work into reusable skills, and has a built-in cron scheduler that delivers results to your chats. It works with any model provider, runs commands locally, in Docker, over SSH or in serverless sandboxes, and has sensible security defaults: the gateway denies unknown users unless you allowlist or pair them, and dangerous commands need approval.

Watch out for: An agent that learns on its own needs supervision. Review its skills and memory now and then.

Choose it if: you want a personal or small-team assistant that is always on and gets better at your work. Setup guide: Hermes Agent on a server.

OpenClaw

An open-source (MIT) assistant, now stewarded by the OpenClaw Foundation, that runs on your own machine and works through the chat apps you already use. It started in late 2025 as Clawdbot, was briefly renamed Moltbot, and became OpenClaw at the end of January 2026, growing into one of the most-starred projects on GitHub.

Why it’s good: Breadth. A central gateway connects the assistant to more than 20 chat channels (WhatsApp, Telegram, Slack, Discord, Signal, iMessage and others), a large ecosystem of skills and plugins shared through ClawHub, and companion apps that add voice, camera, screen and device actions. It works with hosted and local models.

Watch out for: Breadth means a large attack surface. Tools run on the host for the main session unless you configure sandboxing, and its README is explicit that inbound messages are untrusted input. Read its security and sandboxing guides and keep it updated.

Choose it if: you want one assistant across many channels and devices, or a big plugin ecosystem.

Hermes vs OpenClaw in one line

Hermes is the better personal agent that grows with you; OpenClaw is the broader multi-channel hub with the larger ecosystem. Hermes can import an OpenClaw setup (hermes import-agent openclaw, after installing its importer plugin) if you want to try switching.

Control surfaces

T3 Code

A free, open-source (MIT) app that drives the coding agents you already have: Claude Code, Codex, Cursor, OpenCode, Pi and more, using your existing subscriptions.

Why it’s good: Every task is a thread that can get its own Git worktree, so several agents work in parallel without colliding. You review diffs per turn or per branch, revert to any message, open pull requests, and answer approval requests, all in one app on desktop, web or phone. Run its server on an always-on machine and you can start work at your desk and check it from your phone.

Watch out for: It is early software (“Expect bugs”), and it is only as good as the agents it drives.

Choose it if: you run more than one agent task at a time. Full guide: T3 Code: how to use it.

How to choose

• Writing code in a repository? Start with Claude Code or Codex. Add OpenCode if you want open source or local models, or Pi if you want to build your own.
• Running several coding tasks at once, or following them from your phone? Put T3 Code on top.
• Want an always-on assistant in your chat apps? Hermes Agent for a personal agent that learns; OpenClaw for the widest channel and plugin support.
• Want any of these always on? Read why run an AI agent on a server first, especially the security checklist.

What we use

At Poligon we build software mainly with Claude Code and Codex, driven through T3 Code, with skills and hooks encoding how we work. For client operations, none of these general-purpose harnesses is the end product. A business process like order entry or invoice matching needs a narrow harness of its own: specific tools, least-privilege access to the client’s systems, human approvals and a full audit trail. The general tools are how we build it; the narrow harness is what we ship.