Grab's Agent Framework LLM-Kit Accelerates AI Agent Production Deployment

Grab's internal Agent Framework, LLM-Kit, has significantly accelerated AI agent production deployment. The framework standardizes over 500 internal agent services, providing a basic agent loop, evaluation, tracing, secret handling, and tool-server connections. Agents discover tools at runtime from over 50 MCP servers and connect to models via a single gateway, allowing new capabilities and providers to arrive through registration or configuration rather than redeployment.

Wiring a new AI agent service into production now takes about an hour, down from two weeks or more. The reduction in time is not due to the agent's reasoning loop but the surrounding infrastructure such as secrets, tracing, service discovery, and evaluation. Grab's LLM-Kit is described in an engineering post, detailing how the reasoning loop took an afternoon, while the production wrapper took two weeks.

The framework provides scaffolding around existing infrastructure, eliminating the need for per-service decisions. Developers can fill in a form and receive a GitLab repository containing a working FastAPI service, along with LangGraph agent modules, OpenTelemetry tracing, Vault secrets, and service discovery, complete with an evaluation endpoint from the first commit.

Tools are fetched at runtime from over 50 registered servers using the Model Context Protocol, and model calls go through the OpenAI-compatible GrabGPT Gateway, which fronts five providers, injecting credentials to avoid hard-coding provider details in application code. This approach allows for flexibility and avoids the risk of vendor lock-in.

Grab also utilizes Palana, an autonomous agent experimentation tool, and a multi-agent engineering support system for analytics, providing identity, secrets, network access, and operational visibility. LLM-Kit's platform layer can now be bought rather than built, such as Amazon Bedrock AgentCore and Google's Agent Runtime, which offer identity, gateway, and observability for agents written in existing frameworks.