Troubleshooting your AI agent: not replying, not running actions, or stuck in a loop

A step-by-step checklist for the most common reasons an AI agent goes silent, ignores a configured tool, or keeps repeating the same reply.

Jul 11, 2026

Watch this step in the video · Troubleshooting

When an AI agent goes quiet, doesn't run the action you described in the prompt, or seems to repeat the same reply over and over, it's almost always one of these causes. Check them in order.

1. The agent doesn't reply at all

  • Agent inactive: in the agent editor's Identity tab, confirm the status toggle is on.
  • Outside service hours: if Service hours in the Behavior tab is set to Business hours or Custom, outside those windows the AI doesn't reply with the model: it sends the Outside service hours message (if you wrote one) and/or hands the conversation to the team. This is on purpose. In the tester you see the out-of-hours notice and can click Test as if within service hours.
  • Daily token budget: not the cause. That editor field no longer stops the agent (it's disabled across the platform).
  • Plan allowance and Omnifox balance used up: every plan has a monthly AI token allowance. When it runs out, the AI keeps going on your Omnifox balance if it has funds and Keep going with your Balance once you pass the cap is on; otherwise the system stops calling the model and escalates the conversation instead of leaving the customer unanswered. Check usage and top up under Settings → Billing.
  • Channel set to "never": if the channel's AI routing strategy is "never", that channel simply never triggers the bot.

2. The agent replies but doesn't run the expected action

  • The tool is toggled off for that agent: every tool (close conversation, assign, etc.) has its own switch under the agent's Custom actions/editor. If it's off, the model never even sees it as available, no matter what the prompt says.
  • The tool isn't included in your plan: booking, moving and cancelling appointments in the Omnifox calendar is on every plan; HTTP actions require Impulsa or above; updating contact fields or tags, triggering a workflow and booking a meeting (Cal.com link) require Crece; voice call handling (handle_call) requires Escala.
  • The prompt doesn't say WHEN to use it: the model only calls a tool when the system prompt explicitly describes the trigger condition (e.g. "if the customer confirms it's resolved, close the conversation").
  • Malformed parameters: if a tool needs an exact value (e.g. a lifecycle stage slug or a numeric workflow_id) and the model doesn't have it, the call fails silently for the customer but shows up in the Sandbox trace/log.

3. The agent gets stuck in a loop / repeats the same reply

  • Use the editor's Sandbox (live tester) to replay the conversation and inspect the full turn-by-turn trace.
  • Turn on the anti-loop lock: under the Behavior tab, "Escalate to a human after N replies without resolving" cuts the cycle automatically after N replies in a row without a resolution (see the dedicated article).
  • Check the confidence threshold (auto_handoff_confidence_threshold): if it's set too low, the model may keep replying even when it isn't confident.
  • If the customer sends several messages in quick succession, anti-spam (Behavior → Anti-spam) batches them into a single reply — if it's off, you may see duplicated or crossed replies.

Example

A Sales agent never closed conversations even when the customer said "thanks, all good". Checking the Actions tab, close_conversation was toggled off for that agent — flipping it on (the prompt already had the instruction) fixed it immediately.

Tips

  • Always test changes in the Sandbox before assuming "the model got it wrong" — most cases are configuration, not the model.
  • If the issue persists after checking all of the above, contact support with the affected conversation ID.
Was this helpful?

Related articles

Message us on WhatsApp