Skip to main content
Most Agent Mode issues fall into one of these categories: plan eligibility, connector authorization, prompt clarity, or unexpected agent behavior. Work through the relevant section below.
If you don’t see the Agent pill in the prompt bar, or it appears grayed out and can’t be clicked:
  1. Check your plan — Agent Mode requires a Pro plan or above. Free plan users do not have access.
    • To upgrade: Settings → Account → Subscription
  2. Refresh the page — if you recently upgraded, a page refresh is needed to activate new plan features.
  3. Log out and back in — plan changes sometimes require a fresh session to take effect.
  4. Check for browser extensions — some ad blockers or privacy extensions can interfere with UI elements. Temporarily disable them and reload ZeroTwo.
If you’re on a Pro plan and the pill is still missing or non-functional, contact ZeroTwo support.
If Agent Mode appears active but ZeroTwo is only responding with text and not taking external actions:
  1. Verify the Agent pill is toggled ON — the pill should be highlighted (blue) before you send your message. If it’s not highlighted, click it to activate, then resend.
  2. Be explicit in your prompt — phrase it as a multi-step task with clear action verbs and named apps: 
    // Unlikely to trigger agent actions
"What would you do to summarize my GitHub issues?"

// Clear agent prompt
"Using GitHub, get all open issues assigned to me and create
a Notion page with a summary table: Title | Repository | URL"
  3. Ensure relevant connectors are connected — if ZeroTwo can’t find a connected app to use, it may fall back to a text-only response. Check Settings → Connectors.
  4. Re-toggle the Agent pill — click it off, then on again, then resend your message.
If Agent Mode reports it can’t access an app you expected it to use:
  1. Check that the app is connected — go to Settings → Connectors and confirm the app shows as Connected with no error indicators.
  2. Check for expired authorization — OAuth tokens expire over time. An expired token will show an error or “Reconnect needed” in the Connectors list.
    • Fix: Click Disconnect next to the app → Connect → complete OAuth authorization again
  3. Check the granted permissions — if the agent needs to write (create issues, send emails) but you only granted read access at connection time, it will fail for write actions.
    • Fix: Disconnect and reconnect, granting the appropriate write permissions during OAuth
  4. Check if access was revoked externally — some apps (GitHub, Google) allow you to revoke ZeroTwo’s access from their own settings pages. If you revoked access there, reconnect from Settings → Connectors.
  5. Check connector-specific scope — some connectors (like GitHub) ask which repositories to grant access to. If the agent is trying to access a repo not included in your original grant, it will fail. Reconnect and expand the scope to include the needed repos.
If Agent Mode is taking actions you didn’t intend:Stop immediately — click the Stop button in the prompt bar to halt all execution.Review what happened:
  1. Open the Agent Activity sidebar — it shows every action taken in order, with timestamps
  2. Check which actions completed before you stopped
  3. If any completed actions need to be undone (e.g., a GitHub issue was created that shouldn’t have been), handle that manually in the app
Refine your prompt before retrying:
  • Add explicit stopping conditions: “Stop after creating the Notion page; do not send anything”
  • Name specific targets: “…in the ‘api-service’ repository only”, “…to the #engineering Slack channel”
  • Add scope limits: “Only read — do not create or modify anything”
Adjust confirmation sensitivity: Go to Settings → Preferences → Agent Confirmations and set it to Ask for all actions so every external action requires your approval.
If confirmations are interrupting your workflow more than you’d like:
  1. Go to Settings → Preferences → Agent Confirmations
  2. Change the setting to Ask for high-impact only — this reduces confirmations to sends, creates, deletes, and irreversible actions while skipping confirmations for routine read operations
Only reduce confirmation frequency after you’ve thoroughly tested your workflows and are confident in the agent’s behavior. Fewer confirmations means Agent Mode can take more actions without pausing for your approval.
If Agent Mode keeps repeating the same step or appears stuck without making progress:
  1. Click the Stop button — this is always the safest first action. It prevents runaway API calls or repeated unintended actions.
  2. Review the Activity sidebar — look for which step the agent is repeating and why. Common causes:
    • A connector is returning an error and the agent is retrying automatically
    • The objective is ambiguous and the agent doesn’t know when it’s done
    • A required resource can’t be found and the agent keeps searching
  3. Restart with a clearer prompt — add explicit success criteria and fallback conditions: 
    // Original (looped)
"Organize my GitHub issues"

// Improved
"In my 'api-service' repo, label all open issues older than 30 days with 'stale'.
Do not close them. Stop after applying labels to all qualifying issues.
If there are no qualifying issues, say so and stop."
  4. Check connector health — if a connector is consistently failing (rate limit hit, expired token), the agent may loop on that step. Reconnect the connector in Settings → Connectors and retry.
If ZeroTwo shows a prompt injection warning in the Agent Activity sidebar:This means ZeroTwo detected content in an external source (a web page, email, document, GitHub issue, etc.) that appears to contain instructions attempting to redirect the agent’s behavior.What to do:
  1. Do not approve the flagged action — click Cancel or Stop All in any pending confirmation dialog
  2. Review the flagged source — the warning indicates which source triggered it (e.g., “Suspicious content detected in GitHub issue #42”). Inspect that content manually.
  3. Determine if it’s a false positive — some legitimate content (e.g., a document about AI prompt techniques) may trigger the detector. If you’re confident the content is safe, you can proceed carefully.
  4. If the content appears malicious — do not process it further. Report the source if appropriate. Contact security@zerotwo.ai if you believe this is a targeted attack.
Never override a prompt injection warning without carefully reviewing the flagged content yourself. These attacks are designed to look legitimate.

​
Getting More Help

If an issue isn’t resolved by the above: