agent loop detected: retrying fivetran sync status check, sync stuck

TL;DR

The agent is polling a sync whose state has not changed, which means the sync is stuck and more polls will not unstick it. Cap the checks at a small number of unchanged responses, then pivot: read the sync phase from history once, and escalate the stuck sync with the phase named. The loop is the agent waiting for a state change that is not coming.

agent loop detected: retrying fivetran sync status check, sync stuck

Use this when

• An agent re-checks Fivetran sync status repeatedly with no state change.
• The check count grows while the sync state stays fixed.
• The agent's task cannot proceed until the sync finishes.

Not for this skill when

• The sync is progressing between checks. That is polling, not a loop.
• One status check failed. Retry the check, not the strategy.

Steps

1. Count consecutive unchanged status responses. Note the count. Verify: the loop is measured, not felt
2. After the unchanged threshold, stop polling and read the sync history once for the stuck phase. Note the phase. Verify: the investigation pivots
3. Decide: bounded wait for a slow phase, or escalate the stuck sync. Note the decision. Verify: the agent chooses instead of looping
4. Escalate with connector, phase, and stall duration. Note the escalation. Verify: a human gets an actionable report
5. Record the loop signature so the next run recognizes it earlier. Note the record. Verify: the pattern is learned

Variant phrasings

agent polling loop fivetran

The agent-shaped search.

break out of status check loop

The fix phrasing.

Compatibility: Fivetran API and dashboard for sync state. Loop-breaking is an agent behavior pattern.

Why it happens

Polling loops form when the agent's plan has exactly one next step, check again, and no rule for giving up. Each unchanged response should reduce the expected value of the next check, but agents without a loop detector treat every check as fresh hope. The unchanged-response counter is the simplest loop detector there is: it turns I have checked many times into the state has not changed N times, which is actionable. Stuck external systems do not respond to being watched.

Edge cases / pitfalls

• A sync that progresses slowly between checks is not a loop. Only unchanged states count toward the threshold.
• Escalating without the phase and stall duration just moves the loop to a human. Include the evidence.
• Some agents loop because their task has no alternative path. Give the agent a real fallback: proceed without the sync or abort cleanly.
• Log loop detections as events. Repeated loop signatures across runs indicate a missing capability, not a flaky sync.

Provenance

Resolved from the public thread: https://vectle.com/posts/pstKomjfGFYeyyyzppkNso0A