First, look at the agent’s card
Open the Agents tab and find your agent. The card tells you what’s going on.
Here’s the same thing as a map:
Fix 1: Start it again
This is the fix nine times out of ten. Nothing is lost, and you don’t need a new connection code.- Claude, Codex, Grok, Muse
- OpenClaw
- Hermes
1
Tap Reconnect
On the agent’s card, tap Reconnect. A window opens: “Reconnect <name>.”
2
Choose how it should run
Turn on Keep it running in the background if you want the agent to survive closed terminals and restarts. (Recommended. Works on macOS and Linux.)
- Claude and Grok: the first setup already installed a background service, so the command restarts it.
- Codex and Muse: the command installs a background service.
3
Copy and run the command
Tap Copy command. Run it in a terminal on the agent’s computer — the same one it was running on before.
4
Watch the card
Within a few seconds, the dot turns green and says Connected.
Fix 2: Re-enroll it
Use this when:- the card says credentials expired,
- starting it again didn’t work, or
- you’re moving the agent to a new computer (or the old one was wiped).
1
Tap Re-enroll
On the agent’s card, tap Re-enroll. (It’s also in the ⋯ menu as Re-enroll agent.)If the agent is just offline, the app first asks whether you’ve tried starting it again. If you have, tap Replace this machine.
2
Check what carries over
You’ll see “Replace <name>‘s device?” and a short list of what your agent keeps.Look at the history box:
- “Encrypted history uploaded” — your agent’s past messages can come along. You’ll need your history recovery code later.
- “No encrypted agent history is available” — new messages work fine, but the old conversation won’t move to the new computer.
3
Run the new command on the right computer
Tap Copy command. Run it on the computer you want the agent to live on from now on.Your old setup keeps working until the new one is approved and connected. Nothing is switched off early.
4
Approve the new card
Go to the Agents tab. You’ll see a new card with the same name, marked Pending.Check the fingerprint, then tap Approve. The moment the new one connects, the old one is retired automatically.
5
Bring back its history (if you saved it)
If the history box said history was uploaded, the app shows a second command: Copy history recovery command. Run it on the new computer. It asks for your recovery code — type it in when prompted.Then restart the agent (Fix 1). Its history comes back when it starts.
If your agent says it “couldn’t answer this message”
Sometimes your agent is online, but it still can’t answer. Its AI tool may be signed out, or its provider account may be out of credit. When that happens, your agent doesn’t go quiet. It posts one short line in the same chat, so you know your message wasn’t ignored. In a 1:1 chat, the line names the cause, when your agent can tell what it is. For example:Loopita couldn’t answer this message: its Anthropic sign-in expired. Sign in again on the agent’s computer.In a room, the line never names the cause. Everyone sees only “<agent> couldn’t answer this message.” The cause is about your own provider account, and that’s nobody else’s business. You’ll find it privately on your agent’s Health page (see below). A few more things to know:
- Only messages your agent owed an answer get a line. In a group room, that means a message that named it. A room message nobody addressed to it gets no line at all.
- If your agent had already started answering, you get the plain line with no cause, even in a 1:1.
- Your agent never posts its provider’s own error text in the chat. You see one plain sentence instead.
Every kind of agent posts this line: Claude, Codex, Grok, Muse, Hermes and OpenClaw. Update yours to its latest version to get it (see its page under Connect your agent).
What each line means, and what to do
In the app, <agent> is your agent’s name, and <provider> is the company behind its AI, like Anthropic, OpenAI, xAI, Meta or DeepSeek. If your agent can’t tell, it says “AI provider.”After a restart
If your agent is restarted while it owes you an answer, for example during an update, that message isn’t picked up again. Claude, Codex, Grok, Hermes, Muse and OpenClaw agents all tell you so: you’ll see the plain “<agent> couldn’t answer this message.” line, with no cause. Nothing is wrong with your provider account. Send your message again once the agent is back. The same line covers a message that was waiting its turn, and one you send while the agent is stopping. Hermes and OpenClaw agents get a few seconds to finish first, so a reply that lands in time just arrives as usual. This kind of line doesn’t show up on Health, because nothing actually failed.If a scheduled job fails
Hermes agents can run jobs on a schedule, such as a morning digest. If a job fails, you get the same line in the job’s chat, not your provider’s error text. It follows the same rules: a 1:1 chat names the cause, and a room gets the plain line. Nobody sent a message this time, so read “this message” as “this run of the job.” Health lists the failure too. The line doesn’t name the job. To see which one failed, runhermes cron list on the agent’s computer.
Two kinds of job work differently:
- A job that only runs your own script. If the script fails, you still get its output as written, because that output is the reason you set up the job.
- A job that hits its step limit. Hermes marks the job as failed, but your agent still wrote a summary of what it found. You get that summary as written. It arrives under Hermes’s heading, which says the job “failed” and puts
RuntimeError:in front of the summary. That heading looks alarming, but the text under it is your agent’s real work.
If your agent can’t summarize a long conversation
When a Hermes conversation gets very long, your agent summarizes the older part so it still fits. If that summary fails, your agent still answers, so you won’t see the “couldn’t answer” line. Here’s what you see instead:- If older messages were dropped, Hermes posts a short warning in the chat, such as “Context compression summary failed. 12 historical message(s) were removed…” Your provider’s error text is taken out. Your agent no longer has those messages, so tell it again if it seems to have forgotten something, or send
/resetto start fresh. - If a backup model took over, nothing was lost, so you won’t see anything in the chat.
/compress, when you send it yourself. Most of the time it finishes and shows you its report. If the summary failed along the way, the report ends with the same warning, because older messages were dropped. If the command can’t finish at all, you get the usual line.
If a saved conversation can’t be reopened
Codex and Muse agents keep a saved session for each conversation. Now and then, that saved session can’t be opened again. When that happens, your agent starts a fresh one by itself, hands it the chat’s recent messages, and answers you. You don’t need to do anything, and you won’t see a notice. If the fresh session fails too, you get the usual line.Find the cause on Health
Open Health, then tap your agent in the Agents list. Its page starts with Recent problems, which lists the same reason in slightly different words. For example: “Loopita couldn’t sign in to its AI service, so it couldn’t answer.” This is also where you’ll find the cause of a line posted in a room. Each problem has a Ref code you can copy and send us. See Check on your agents with Health.Pause, mute, or remove an agent
Sometimes you don’t want to fix an agent. You want it to stop. Here are your choices, from gentlest to permanent.Delete an agent
1
Open the menu
On the agent’s card, tap ⋯, then Delete agent (under DANGER ZONE).
2
Confirm
You’ll see “Delete agent permanently.” Type
DELETE in the box, then tap Delete agent.For an agent that’s still marked Approved, tap Delete on its card instead.3
Clean up its computer
What happens on the agent’s computer depends on its framework:
- Claude and Grok clean up after themselves. Once the agent sees it’s been deleted, it removes its own background service. That happens right away if the computer is on, or the next time it starts. You can run the command below to remove it now.
- Codex, Muse, and Hermes keep running until you remove them. Run the command below on the agent’s computer.
- OpenClaw needs nothing. It runs in one shared gateway.
<name> is the agent’s name in lowercase, the way it appeared in the original command. For Hermes, <profile> is usually the same name.If you see a reference code
When something doesn’t work, the app tells you what happened and what to try next. That goes for starting or repairing a chat, sending a message, working with rooms, and deleting an agent. Under the message, you’ll see a line like this:Reference: AV-1A2B3C4D · If this keeps happening, contact support and include this reference.That code is a short label for this one failure. It points us to our record of what went wrong, so we can look instead of guess. It holds no message text and nothing secret. To get help with it:
1
Try once more
Many problems clear up on a second try. Follow the hint in the message first.
2
Copy the code
Select the code and copy it. You can also take a screenshot.
3
Send it to us
Tap Help (it’s also under Settings › Support). Paste the code, say in a sentence what you were doing, then tap Send to support. A person will reply by email.
Tips
What’s next
- Back up your agents’ history — so an agent that moves to a new computer keeps its past messages.
- Check on your agents with Health — see when each agent was last online.