The model produces text and decisions; the engine runs the loop that turns those decisions into actions and feeds the results back to the model.

What does an agent engine do?

An engine typically handles:

  • Memory: saving what matters and loading it into later conversations.
  • Tools: connecting the model to files, messaging and other software, often through the Model Context Protocol.
  • Schedules: running tasks at set times without you prompting it.
  • Skills: storing methods the agent can reuse, covered under agent skill.
  • Control: approval steps, permissions and a log of what the agent did.

Without an engine, a model can describe how a job should be done but has no means of doing it, much like a carpenter with no workshop.

Which agent engines are there?

OpenClaw and Hermes Agent are two well-known open-source engines. Both can run the same underlying models, and they behave differently because of the software around the model: Hermes keeps a deliberately small, capped memory, for example, while OpenClaw keeps fuller notes it can search. Our OpenClaw vs Hermes comparison goes into the detail. NanoClaw is another open-source engine.

Why does the engine matter so much?

It explains why two products running the same model can feel nothing alike. One has persistent memory and scheduled tasks; the other is a chat window with a system prompt. "Which model does it use?" only covers half of what you're choosing, and the other half is what the engine adds.

The engine is also the longer-term decision. You can swap the model inside an engine and keep the agent's memory, tools and skills. Swap the engine and you get a different agent even if the model stays the same, because the engine is where your setup lives.

An example

You send two assistants the same message: "Every weekday at 8am, send me the three things I said I'd finish today."

An agent whose engine supports scheduled tasks creates a recurring job and messages you at 8am each weekday. A plain chat interface explains how you could set up a reminder in another app, because it has nothing that can run on a timer. Both models understood the request, and only one was connected to something that could carry it out.

Common questions

Is an agent engine the same as an agent framework?

Broadly, yes. "Framework" is more common among developers and usually means a code library you build with. "Engine" or "harness" tends to describe the running software that a finished agent sits on.

Do I need to choose an engine?

Only if you're running an agent yourself, or if a hosted product offers a choice. Otherwise the provider picks one for you. Able°, for example, runs on Hermes by default, with NanoClaw as the alternative. Whichever it is, judge it by what it gives you in practice: memory, schedules, skills and control.

Can I move my agent from one engine to another?

Sometimes, with some effort. Memory and skills are stored in formats that don't always transfer cleanly, although some engines include import tools, such as Hermes Agent's importer for OpenClaw setups. Ask how export works before you commit.