Hermes Agent vs OpenClaw: why JB switched
This comparison of Hermes Agent and OpenClaw comes from running a real business on both, not from a test project. JB started on OpenClaw in early 2026, coordinating several agents for a law firm Jon owns. By July 1, 2026, all of JB’s agent work ran on Hermes and the OpenClaw gateway was switched off.
Both projects move fast, and some of what follows may already be fixed. Read it as one team’s experience in the first half of 2026, not a verdict.
What OpenClaw did well
OpenClaw got us started. One config file described every agent, which model each one used, and which chat channel it answered on, so we could run a coordinator on a stronger model and cheaper models for routine work. Telegram and Signal bots came up quickly, and scheduled heartbeats kept the work moving between conversations.
What broke for us
- Scheduled jobs fired hundreds of times. On the version we ran in February 2026, each cron job fired more than 200 times in about a second every time it was due, because the next run time did not advance. We traced it and wrote up a bug report.
- The central config got overwritten. Over two days in April, the main config file was overwritten three times. With every agent in one file, one bad write touched everything.
- Agents pretended to use tools. When an agent ran on a model route without tool support, or with a tool profile that sent no tool definitions, it wrote out fake tool calls and fake results instead of failing. That was partly our configuration, but it was hard to notice and costly to trust.
Why Hermes fits JB now
Hermes was already running JB’s work when, in September 2026, JB became a product: one agent per business, reached by text, with approvals that hold. Its shape fit that better than anything we had. Hermes profiles map one to one onto businesses, its BlueBubbles adapter put JB in iMessage, commands run in a sandbox per profile, and routines can run plain scripts on a schedule without a model call.
Side by side
| OpenClaw (as we ran it) | Hermes Agent (as we run it) | |
|---|---|---|
| How agents are organized | Many agents in one central config, each bound to a channel | Profiles, each with its own instructions, memory, and tools |
| One agent per customer | Possible, but every agent shares the one config | A profile per customer is the natural unit |
| iMessage | Not something we used it for | Built-in BlueBubbles adapter |
| Scheduled jobs | Built-in cron, which misfired on the version we ran | Routines, including script-only ones that never call a model |
| Running commands | On the host in our setup | Sandboxed terminal per profile |
Which should you pick?
If you want several agents for yourself, wired to the chat apps you already use, OpenClaw is a quick start. If you are building one agent per customer, need iMessage, or want each agent boxed in with its own tools and sandbox, Hermes is the better fit. Either way, test the scheduler and the tool calls before you trust them with anything that matters.
And if you would rather not run any of it yourself, that is what JB is for.
Next: how JB texts over iMessage on Hermes.
Want JB on your business?
JB looks up your business and shows you three fixes, free. See what JB would fix for you.