Safety & enforcement

Loop detection

Zelyx watches for repeating call patterns and can warn or block when an agent appears to be looping. The single biggest cause of unexpected AI spend is an agent that retries in a tight loop — loop detection catches this before the bill does.

How it works

For each call, Zelyx creates a fingerprint based on the session ID, model, and whether the call uses tools. It then counts how many calls with the same fingerprint have occurred within a rolling time window.

If the count reaches the configured threshold, loop detection triggers and takes the configured action (warn or block).

Configuration

Configure loop detection in Settings → Loop detection.

Actions

Handling a 429 in your agent

import time, httpx

for attempt in range(max_attempts):
    try:
        response = client.chat.completions.create(
            model="gpt-4o-mini",
            messages=messages,
            extra_headers={"X-Zelyx-Session-Id": session_id},
        )
        break
    except httpx.HTTPStatusError as e:
        if e.response.status_code == 429:
            # Loop detected — back off and re-evaluate
            time.sleep(10)
            messages = trim_messages(messages)   # reduce context, try again
        else:
            raise

Resetting a session counter

Loop counters are in-memory and reset automatically on server restart. If a legitimate burst of calls (e.g. parallel document processing) triggers loop detection, the counter resets by itself within the window — no manual action required. If you need an immediate reset, use a different session ID for the new job.