Grab says engineers can now put a new internal AI-agent service into production in about an hour, down from two weeks or more. The change comes from LLM-Kit, a shared framework now used by more than 500 services across the ride-hailing and delivery company.
LLM-Kit does not replace the agent’s reasoning loop. It generates a FastAPI service with LangGraph modules, OpenTelemetry tracing, Vault-based secret handling, service discovery and an evaluation endpoint already connected. Grab’s account is that the distinctive agent logic often took an afternoon, while the production wrapper previously consumed weeks.
Agents discover tools at runtime from more than 50 servers using the Model Context Protocol, a standard for connecting models to external tools. Model calls travel through an OpenAI-compatible gateway that fronts five providers and supplies credentials, so application code does not embed provider-specific secrets.
Centralizing that infrastructure also changes how capabilities are updated: registering a tool or configuring a model provider can make it available without redeploying every agent. The tradeoff is organizational rather than purely technical. Grab deliberately chose a framework over a rigid platform, but teams still share its runtime conventions, evaluation system and gateway. The project shows that production bottlenecks often sit in security, observability and integration rather than the agent loop itself.