Glitch Grow AI Ads Agent: Autonomous Operation Agent for E-commerce Ad Delivery

Glitch Grow AI Ads Agent is an AI ad operation agent designed for Shopify and e-commerce brands, enabling end-to-end management of multi-channel ads (Meta, Amazon, etc.). It implements a complete closed loop of plan-analysis-execution-learning via LangGraph, and integrates a Telegram human-AI collaboration interface to free operators from tedious daily tasks. Its core value lies in turning operators from executors into supervisors by handling autonomous ad operations while allowing human oversight at key decision points.

In digital marketing, ad operation is often tedious and highly repetitive. Operators need to monitor multiple platforms, process massive data, and handle tasks like data analysis, budget adjustment, creative testing, and performance tracking. Glitch Grow AI Ads Agent addresses this pain point by acting as an autonomous agent that can run paid media operations independently, allowing operators to shift from executors to supervisors.

The agent operates on a closed-loop workflow: plan (decide next tests based on history and trends) → analyze (evaluate action effects) → execute (adjust ads/budget per brand thresholds) → learn (accumulate experience into memory). It uses LangGraph as the orchestration framework, which offers persistent checkpoints (for approval waits), model selection per node (cost-cognitive demand matching), deterministic retries (reliability before budget changes), and conditional entry points (supports 12+ command types). The agent also uses an MCP (Model Context Protocol) architecture to decouple data layers (e.g., Meta/Amazon data via dedicated MCP servers) and a memory system (pgvector for retrieval, nightly cron to refine long-term memory).

The agent integrates data from three core platforms: Shopify (GMV, AOV, repeat users, UTM coverage), Meta Ads (campaign/group/ad-level spend, creative, target URL), and Amazon (Seller Central orders, SP ad performance, ASIN profit data). For attribution: Shopify-Meta uses PostHog's ground truth data (instead of Meta's potentially inflated numbers); Meta-Amazon uses a subtraction model (total Amazon orders minus Amazon SP ad orders) when the Amazon Attribution API is unavailable. This enables accurate answers to key questions like real mixed ROAS across channels and ASIN performance issues.

The agent interacts with operators via a Telegram Bot (each workspace has an independent instance, with admin whitelist for security). Current commands include /insights (GMV, orders, AOV), /roas (real vs Meta ROAS), /tracking_audit (pixel/CAPI gap fixes), /ads (ad leaderboard by spend), /creative (Gemini visual assessment), /ideas (creative briefs from winning patterns), /alerts (anomaly monitoring like CPC drift), /amazon (Amazon data summary), /attribution (Meta-Amazon analysis). The upcoming v2 version will add autonomous actions (pause ads, adjust budget) that enter an approval queue when exceeding brand thresholds, allowing operators to approve/reject with one click.

Deployment: Cloud Run hosts LangGraph endpoints and Cloud Scheduler jobs; VM/systemd hosts the Telegram Bot and Shopify webhook receiver (needs local access to meta-ads-mcp). Security: Telegram webhook uses secret_token verification; Shopify webhook uses HMAC verification; /agent/run endpoint requires Bearer token and Cloud Run IAM in production. It is recommended not to use --allow-unauthenticated for any endpoint, and to place Cloud Run behind an HTTPS load balancer with path-based policies.

The agent is ideal for: e-commerce brands managing multiple Shopify stores (need unified multi-channel view), teams shifting from manual reports to automated decisions, brands needing cross-platform attribution for budget optimization, and advertisers wanting to reduce monitoring time while retaining control. Its value lies in freeing operators from repetitive tasks (data checking, spreadsheet work, budget adjustments) to focus on strategy, key decision approval, and optimizing agent rules. The agent complements human judgment instead of replacing it.