Agent not responding

Symptoms#

You send a message in the chat panel and get no response, see an error, or the chat is disabled.

"Agent not ready" message#

Your desktop is still booting. The AI agent is the last service to come online.

Why this happens#

The boot process has several phases. Your desktop may show running and even have a working VNC display, but the agent is not available until it finishes initializing. This takes about 60 seconds after the VM is created, well after VNC and the terminal are up.

The initialization process differs by agent type, but all three (OpenClaw, Claude Code, and Hermes) need to fully start before chat is available.

What to try#

Wait. This is normal during the first 60 seconds. The chat panel enables automatically once the agent is ready. If it has been more than two minutes, stop and restart the desktop.

"Gateway agent failed; falling back" error#

The agent process either crashed or never started.

Possible causes#

The most common cause is an invalid API key. If the key you provided during desktop creation is invalid, expired, or lacks the right permissions, the agent cannot authenticate with your AI provider.

If you selected a model that is not available on your provider, or the model ID is wrong, the agent cannot start a conversation either.

Provider outages (Anthropic, OpenAI, OpenRouter) will also cause this.

What to try#

1. Verify your API key on your provider's dashboard. Confirm it is active and has credits.
2. Check the provider's status page for ongoing incidents.
3. Try the default model to rule out model availability problems.
4. If the API key was wrong at creation time, delete the desktop and create a new one with a valid key. Provider configuration is injected at boot and cannot be changed on a running desktop.

Chat messages get no response#

You send a message, it shows up in the chat, but nothing comes back. No error, no response.

Possible causes#

If you sent many messages quickly, the provider may be rate-limiting you. This is common with free-tier API keys.

Some tasks, especially file operations, web browsing, or complex reasoning, can take 30 seconds or more. The agent may still be working.

The agent process may also have crashed silently after starting. This can happen if the VM runs out of memory.

What to try#

1. Wait 30 seconds. The agent may still be processing.
2. Send a short message like "hello" to check if the agent responds.
3. Open the terminal panel and look for error output from the agent process.
4. Stop and start the desktop to get a fresh agent process.

Troubleshooting by agent type#

OpenClaw#

• If the agent behaves unexpectedly, the soul file may contain instructions that conflict with your request. You can review and edit it from the terminal.
• Verify your API key is valid and has credits on your provider's dashboard.

Claude Code#

• If the agent is unresponsive, check that your Anthropic API key is valid and has credits.
• Claude Code requires an Anthropic API key specifically -- it does not work with OpenAI or OpenRouter.

Hermes#

• If the agent is unresponsive, check that your API key is valid and has credits.
• Hermes is self-improving and may take longer on initial tasks as it builds context.

Provider configuration required#

This means no API key was set when the desktop was created.

What to try#

You cannot add an API key to a running desktop. Delete it and create a new one with:

1. An AI provider selected (Anthropic, OpenAI, or OpenRouter)
2. A valid API key
3. A model chosen

All three are required.

Avoiding agent problems#

• Test your API key on the provider's website before using it in Le Bureau
• Keep your key funded. Most agent failures come down to expired credits.
• Start with a known-good model (Claude Sonnet, GPT-4o) before trying specialized ones
• Give the desktop 60 seconds after creation before expecting chat to work