Grab's Analytics Data Warehouse team built a multi-agent AI system to automate repetitive engineering support tasks across a platform serving 1,000+ internal users and 15,000+ tables. The architecture separates requests into two workflows: investigation (query analysis, log retrieval, schema lookup) and enhancement (code changes, SQL fixes, automated merge requests). Orchestrated via LangGraph and FastAPI, specialized agents handle constrained tasks to improve predictability. The team reduced their internal tool ecosystem from 30+ to a curated set to improve maintainability. Safety measures include SQL validation layers, sensitive data handling, and mandatory human-in-the-loop review for all code changes. Context management across multi-step reasoning is handled through structured compression and selective retrieval. The result is reduced time on routine support and a shift toward higher-value platform engineering work.
Sort: