Start with what an agent actually does
A chatbot answers from what it knows. An agent has access to data (your files, your inbox, a service’s API) and takes actions based on what it finds. So the interesting question about an agent is never which model it runs on. It’s what you’ve allowed it to touch.
So a good first agent is something boring that you already do often enough to write down as a list of steps.
The harness matters more than the model
Every agent needs a layer around the model: tool access, memory, scheduling, credentials, permissions, and somewhere to actually run. That layer is the harness, and it’s where agents succeed or fail. Cowork, OpenClaw and Hermes are all harnesses. They differ in how much they assemble for you, and how much control they hand back.
Claude Cowork: the one you probably already have
Cowork is Anthropic’s agent layer, in the same app as Claude chat. It comes with any paid plan and runs in the macOS and Windows desktop apps, with web and mobile added in July. Point it at folders and connected services and it reads, edits and creates files, and runs tasks on a schedule or a trigger: a summary of what landed overnight, a review that fires when a document changes.
For a first agent the useful part is that you design it by talking to it. Describe the task, let it propose the steps, correct it, run it.
OpenClaw: the always-on connector
OpenClaw is the self-hosted, model-agnostic harness that grew out of Clawdbot and became one of the fastest-growing open-source projects on record. It runs on your own hardware, acts on a schedule rather than waiting for prompts, and connects to the messaging apps you already use: WhatsApp, Telegram, Slack, Discord, Signal, iMessage. Data stays local, and you can point it at Claude, a GPT model or something under Ollama. If you want one agent reachable from your phone that touches everything, this is it.
Hermes: the agent that accumulates
Hermes, from Nous Research, is the other open-source harness people land on. Same self-hosted premise, opposite emphasis. Where OpenClaw is broad and comparatively stateless, Hermes is built around persistent memory: one agent that learns your context and gets more useful the longer it runs. Pick it when the job is automation that should improve over months, rather than orchestration across many services.
Why not just build it with Claude Code?
Claude and Claude Code solve different halves of the problem. Claude chat is for thinking with the model. Cowork is for working alongside you, in your files, with your tools. Claude Code is for building software, and that software runs on its own, as Python or whatever else, with no model inside it unless you write the API calls yourself.
So an agent built in Claude Code means building the harness too: memory, scheduling, retries, permission checks, credentials. Claude Code lets you build anything, which also means you have to build everything. Eventually that’s worth doing, for an agent that looks like nothing off the shelf. For a first one it isn’t.
Which should you start with?
The same comparison as a table
| Setup | Runs | Model | Best for | |
|---|---|---|---|---|
| Claude Cowork | None, it’s in the app | Your desktop, on a schedule | Claude only | A first agent, working today |
| OpenClaw | Self-hosted | Always on, your hardware | Any: Claude, GPT, local | Breadth: many services, reachable from chat apps |
| Hermes | Self-hosted | Always on, your hardware | Any, local-friendly | One agent that accumulates memory |
| Claude Code | You build it | Wherever you deploy it | Whatever you wire in | Bespoke agents, once you know the shape |
If you already pay for Claude and you’re comfortable granting it access to local files, start with Cowork and get something working today. Move to OpenClaw or Hermes when you hit a real limit, usually the agent needing to run without your laptop open, or to be reachable from your phone. Reach for Claude Code when no harness fits.
A first agent worth building
Start with your inbox. You do it every day, you already know what a good answer looks like, and if the agent gets it wrong you’ve just got a summary you ignore. Have it read the overnight mail and tell you what needs attention, what can be ignored, and what you should probably unsubscribe from. Run it read-only for a week and check its judgement against your own.
Then promote it, one action at a time. Drafting replies for your approval. Filing. Unsubscribing. Every action you add is permission you can’t easily take back, so add them slowly and keep it away from anything financial or medical.
If you want the model to teach you the ground rules first, ask it directly and cap the length: “I want to build my first agent. Explain it like I’m ten, in ten steps.” Without the cap you get a lecture instead of a plan.
Frequently asked questions
What is an AI agent?
An agent is a doer rather than a talker. A chatbot answers from what it knows. An agent has access to data such as your files, inbox, calendar or a service's API, and takes actions based on what it finds there. The difference that matters is permission plus the ability to act, not intelligence.
Do I need to know how to code to build an agent?
No, not for a first one. Claude Cowork lets you design an agent in conversation: you describe the task, it proposes the steps, you refine them. OpenClaw and Hermes are self-hosted and expect you to be comfortable installing software, editing config files and managing credentials, but neither requires you to write the agent logic yourself.
What is an agent harness or framework?
The layer around the model that turns it into a working agent: tool access, memory, scheduling, permissions and the execution environment. The model supplies the reasoning and the harness supplies everything else. Cowork, OpenClaw and Hermes are all harnesses. They differ in how much they assemble for you and how much control they hand back.
Should I start with Claude Cowork, OpenClaw or Hermes?
Start with Cowork if you already pay for Claude and want a first agent working today. Choose OpenClaw if you want an always-on agent reachable from your messaging apps and connected to many services. Choose Hermes if you want one agent that accumulates memory and improves over months. Cowork is the lowest-effort start and the easiest to outgrow.
Why not just build an agent with Claude Code?
You can, and for a bespoke agent you eventually will. But Claude Code builds software that runs on its own. If you want a model inside that software you have to write the API calls, the memory, the retries and the permission checks yourself. A harness has those assembled already. Claude Code lets you build anything, which also means you have to build everything.
Is it safe to give an agent access to my email and files?
Treat it as a real risk, not a formality. An agent that reads untrusted content and can also act on your behalf is exploitable through prompt injection, where instructions are hidden in a page or an email it processes. OpenClaw has had several documented vulnerabilities of exactly this kind. Start read-only, grant one action at a time, keep the agent away from banking and health accounts, and prefer setups where you approve actions before they run.