Skip to main content
If the agent isn’t behaving according to your system prompt:
  1. Tighten the language — vague instructions produce inconsistent behavior. Replace soft suggestions with explicit directives: 
    // Vague
"Try to be concise"

// Clear
"Limit all responses to under 200 words. If a topic requires more detail,
offer to expand on specific points."
  2. Use a numbered list for critical rules — bullet lists and numbered rules are more reliably followed than prose instructions.
  3. Remove ambiguity — if two parts of your system prompt could conflict, clarify the priority. Agents follow ambiguous instructions inconsistently.
  4. Add an example output — include a short example of the ideal response format at the end of your system prompt. This shows the agent exactly what you expect.
  5. Test with a fresh chat — system prompt changes take effect in new chats, not ongoing ones. Start a new chat with the agent after making changes.
If the agent isn’t using the tools you expect, or is using ones it shouldn’t:
  1. Verify tool configuration — open the agent editor (/agents → select agent → Edit) and check which tools are toggled on. Make sure the tools you want are enabled and unwanted ones are disabled.
  2. Re-save after changes — toggle tools on/off → click Save. Changes don’t persist until saved.
  3. Be explicit in your prompt — mention the tool in your message: “Use web search to find the latest…” or “Create a canvas document with…”
  4. Check if the model supports the tool — some tools may not be available with every base model. If a tool toggle appears grayed out, try switching to a different base model.
If the app you want doesn’t appear in the agent’s connector list:
  • The app must be connected globally first — go to Settings → Connectors and connect the app there. Once connected at the account level, it appears as an option in the agent’s connector settings.
  • After connecting the app in Settings, return to the agent editor and you should see it available to select.
If the app is already connected globally but still doesn’t appear in the agent editor:
  1. Refresh the page
  2. Open the agent editor again
  3. If still missing, try disconnecting and reconnecting the app in Settings → Connectors
Agent chats are stored the same way as regular chats. If you can’t find a previous chat with an agent:
  1. Check Chat History — go to /c or open the chat history sidebar. Agent chats appear there, usually with the agent’s name or the first message as the chat title.
  2. Search chat history — use the search bar in the sidebar to search for keywords from your previous chat with the agent.
  3. Look for the agent’s name — chats with custom agents are often labeled with the agent’s name. Look for it in the chat list.
If you’ve confirmed the chat is missing (not just hidden), contact ZeroTwo support — this may indicate a sync issue.
If you’ve hit an agent creation limit:
  1. Check your plan — different plans may have limits on the number of custom agents. Check your current plan at Settings → Account → Subscription.
  2. Delete unused agents — if you have agents you no longer use, delete them to free up slots. On the /agents page, find the agent → open its settings → Delete.
  3. Contact support — if you’re on a plan that should include unlimited agents and you’re still hitting a limit, contact ZeroTwo support.
If the agent behaves differently from chat to chat even with the same prompt:
  • AI responses are inherently probabilistic — some variation is normal. The same input doesn’t always produce identical output.
  • To improve consistency:
    1. Add explicit output format instructions to the system prompt: “Always structure responses as: [Section 1] [Section 2]”
    2. Use a less creative / lower temperature model — reasoning-focused models tend to produce more consistent outputs than generative ones
    3. Add few-shot examples in the system prompt — showing 1–2 examples of ideal input/output pairs significantly improves format consistency
    4. Use more specific prompts in chat — vague questions invite more varied interpretations
  • Check for system prompt conflicts — if parts of your system prompt can be interpreted multiple ways, the agent may follow different interpretations each time. Clarify or remove ambiguous instructions.

​
Getting More Help

If an issue isn’t resolved by the above:
  • Review Configure an Agent for guidance on editing models, prompts, and tools
  • Review Agent Tools and Connectors for permission and connector issues
  • Contact ZeroTwo support with a description of the agent’s behavior, your system prompt (redacted if needed), and examples of the unexpected output