Agent to Validate Results

We use LLMs to validate the web-search results produced by the main LLM agent orchestration — a second layer of automated checks, anomaly detection, and self-healing that keeps hundreds of per-site agents honest without constant human babysitting.

Architecture

SQL Tables every consumer below reads from all three AppLog centralised error log for the whole platform flag: is_web_search QaWebSearchMonitoring per-run QA snapshots & summary metrics coverage, drift, field checks PageAccessed audit trail of every page the agents visited URL, status, timestamp Cron Job scheduled aggregation over all 3 tables computes daily/weekly summaries → writes results to a Google Sheet for at-a-glance human review Simple Anomaly Detection deterministic thresholds & heuristics reads all 3 tables · no LLM cost → sends email when something looks off (spike in errors, coverage drop, etc.) Validation LLM scans all 3 tables for fixable patterns picks good candidates → fixes them self-healing loop over the agent fleet Google Sheet human-readable dashboard Email Alert on-demand notification Code / Config Fix applied back to the codebase How the fix is applied Today: a template prompt the operator runs. Possible: swap the operator for an API key → the CI/CD pipeline would auto-deploy the fix. We keep the human in the loop purely to save money on tokens.