My weekly content agent does two jobs: publish one technical article, then find comments I have not answered.
The publishing path was one API call. The engagement path quietly became a fan-out.
It fetched metrics for my latest 20 articles, then requested the full comment tree for every article with comments. On this run, the sixth comment request returned 429 Too Many Requests.
That is not just a rate-limit problem. It is a correctness problem.
If an agent turns a failed request into an empty list, “I could not read the comments” becomes “there are no comments.” The report looks clean while the work is incomplete.
I changed the workflow to be deliberately bounded:
articles = get_recent_articles(limit=20)
engagement_window = articles[:10]
for article in engagement_window:
if article["comments_count"] == 0:
continue
comments = get_comments(article["id"])
# 429 or partial data fails the scan; it never means []
The agent now:
- reads article metrics once
- checks only the 10 newest articles for engagement
- fetches a thread only when
comments_countis non-zero - records
429or partial data as an incomplete scan, not zero comments - keeps article publishing independent from comment-reply availability
The bounded scan needed three comment-tree requests and found three unanswered comments across two recent articles.
Each proposed reply then went through a separate review for self-replies, duplicate replies, wrong parents and generic wording. All three passed.
Posting still failed, correctly. The authenticated web route returned:
{"error":"Please sign in"}
So the agent posted nothing and kept the reviewed replies as pending. No fake success, no blind retry, no comment attached to the wrong thread.
The metrics were useful too. The previous agent-infrastructure post had 23 views and two detailed comments, while several recent posts had no discussion. That is a better topic signal than guessing from titles.
The practical lesson is boring: autonomous agents need bounded reads and explicit unknown states. Rate limits, logged-out sessions and partial API responses are normal operating conditions.
An empty result is data. A failed result is not.
Top comments (0)