This article describes how Grab's internal AI platform SpellVault evolved from a no-code LLM app builder into an agentic AI platform capable of reasoning and acting dynamically.
- •SpellVault started as a RAG-based no-code app builder with plugin support, enabling thousands of internal apps for productivity and automation.
- •The platform expanded with integrations for Slack and Kibana, citation capabilities, Python runtime for math problems, auto-updating Knowledge Vaults, and a Task Scheduler feature.
- •A drag-and-drop Workflow interface was added to let users design deterministic workflows combining LLM calls, Python execution, and Knowledge Vault lookups.
- •The execution model shifted from a legacy executor to a graph-based executor supporting branching, looping, and ReAct agent behavior for all existing apps.
- •Deep Research capability was introduced enabling search across internal repositories (Slack, Wiki, Jira) and the web, with internal features repositioned as autonomous LLM-us