Grab's Agent Framework LLM-Kit Accelerates AI Agent Production Deployment
| Source: InfoQ AI/ML
Tags: Grab, LLM-Kit, MCP, agent-framework, LangGraph, production-AI
Grab cut AI agent deployment time from 2 weeks to 1 hour by building LLM-Kit, an internal framework now backing 500+ services — the savings come from centralizing secrets, tracing, evaluation, and tool discovery, not from the reasoning loop itself.
Details
Grab's engineering team published details on LLM-Kit, the internal framework underpinning more than 500 agent services at the Southeast Asian super-app. Before LLM-Kit, each new agent service required two weeks or more of scaffolding work — wiring Vault secrets, OpenTelemetry tracing, service discovery, and evaluation endpoints from scratch. LLM-Kit handles all of that once, centrally, so a new service is provisioned from a form submission into a working GitLab repository with LangGraph agent modules already configured. Tools are not pre-bundled: agents fetch capabilities at runtime from over 50 registered MCP servers, meaning a new tool reaches every agent without redeployment. All model calls route through GrabGPT Gateway, an OpenAI-compatible proxy fronting five providers, keeping credentials out of application code. Grab deliberately chose a framework over a platform to avoid locking teams into rigid assumptions that age quickly. The evaluation infrastructure is notable: every service gets scoring endpoints from the first commit, using ROUGE, BLEU, and a second model as grader. This treats agent evaluation as a first-class operational concern. The pattern — centralize infrastructure, leave agent logic to teams — is replicable at any company operating more than a handful of agent services.