AI BriefingAI Agents That Self-Recover from Failures
Key point
Socar's Modu-ui-Jucharjang built a LangGraph-based AI agent system that automatically recovers from discount-application failures across 50+ external equipment vendor websites.
Details
Modu-ui-Jucharjang operates a discount application batch that automatically logs into the admin web pages of 50+ equipment vendors to apply discount coupons for on-site exit after parking ticket purchase. External web pages change their UI without notice, causing rule-based automation to break frequently, resulting in approximately 300,000 failures and 165,000 refunds in 2025 alone, with an average of 4 hours needed to respond to each failure.
To solve this, the team fixed the overall flow using a LangGraph StateGraph, designing it so that the LLM only makes decisions within each node. For browser manipulation, they used Playwright MCP to provide the LLM with structured snapshots based on the accessibility tree, and adopted Gemini 2.5 Pro as the model. The LLM cost per recovery is in the range of a few hundred won.
The criterion for splitting agents was changed from 'by failure cause (DOM/Network/Policy)' to 'the order in which developers actually respond.'
- login — perform login
- search — look up vehicle
- discount — apply discount coupon
- check — verify application
- compare — compare screen/API changes and generate a PR
The cause-based design had a problem where, if the routing classifier misjudged, the wrong agent would take over the analysis, and it also did not match how actual operators respond.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.