Build a Hermes agent

Every other runtime on this site answers messages. Hermes runs processes — multi-step work with hand-offs between agents, a queue in front of it, and a place for a human to say yes before something goes out.

That distinction is the whole decision. If you find yourself trying to make a personal assistant handle a pipeline, you want this instead.

Workflows, not conversations

The natural way to build "triage this, research it, draft a reply, let me approve it" on a chat runtime is one very long prompt and a lot of hope. It half-works, and it fails in the specific way long prompts fail: step three forgets what step one decided, and you cannot tell which step was wrong because they were never separate things.

Hermes models those as discrete steps with explicit hand-offs. Each step has its own job and its own context. Practically:

• You can see where it failed. A workflow that dies at the research step is a debuggable event. A long prompt that produced a bad answer is a mystery.
• You can put a human in the middle. An approval step is a step, not a convention you hope the model honours.
• Steps can use different models. Triage is a cheap classification. Drafting is not. Paying drafting prices for triage is a very common and very avoidable way to overspend.

The queue is doing real work

Queue-based execution sounds like plumbing until the first burst arrives. Twenty requests landing in a minute against a chat-shaped agent means twenty concurrent model calls, rate limits, and dropped work. Against Hermes it means a queue that drains at whatever rate the provider allows, in order, with nothing lost.

If the agent is answering for a team rather than for you personally, this is not optional. Teams produce bursts.

Building one

1. Describe the process as steps. Write the brief the way you would explain the job to a new hire: what arrives, what happens to it, who checks it, what leaves. The step boundaries in your description become the step boundaries in the workflow.
2. Be explicit about the approval gate. Say exactly what the agent may send on its own and what needs a human. This is the single most important line in the brief for anything customer-facing.
3. Name the inputs and outputs. Where does work arrive — email, Slack, a form? Where does the result go? Hermes speaks the channels in the fact card on both ends.
4. Deploy and run it on real work with the gate closed. Let it draft everything and send nothing for a few days. You will learn more about the brief in that week than in any amount of testing.

Where the value shows up

The pattern that consistently works is triage → research → draft → human approval, wrapped around a queue that a team already feeds: a shared inbox, a support channel, an intake form. The agent does the reading and the first draft, which is most of the time; the human does the judgment, which is most of the value.

What does not work is pointing Hermes at a job with one step. The orchestration is overhead if there is nothing to orchestrate — that is a OpenClaw or PicoClaw job.

Cost behaves differently here

Multi-step means multiple model calls per unit of work, so cost per item processed is higher than a single-turn agent — and cost per useful outcome is usually much lower, because the alternative is a person doing the reading. The number to watch is cost per completed workflow, not cost per call, and it is on the Billing tab for exactly that reason.

The failure mode to watch for is a step that retries: a research step that cannot find what it needs and keeps trying is the expensive way for a workflow to fail. Give steps a clear "give up and escalate" instruction in the brief.

Hermes inside a NemoClaw box

If you need several workflows isolated from each other — one per team, one per client — NemoClaw can run Hermes inside each of its sandboxes on one dedicated box. That is the shape to reach for when the requirement is "these workflows must not be able to see each other."

Is Hermes the right pick?

Not sure? Describe the job rather than the runtime — the builder's solutions engineer reads the same capability matrix this page does and will argue for a different one if it fits better.