AI
Grab slashes AI agent deployment from weeks to hours with standardized framework
The Southeast Asian tech company's internal LLM-Kit framework has cut production setup time dramatically by centralizing infrastructure, secrets, tracing, and tool discovery across 500+ agent services.
InfoQ
Grab has eliminated months of production infrastructure work from its AI agent rollout cycle. The company's LLM-Kit framework—now standardizing more than 500 internal agent services—brings agent-based features to production in roughly one hour, compared to the two weeks or longer it previously took.
What changed#
The breakthrough wasn't in agent reasoning itself. Rather, Grab moved from solving the same non-reasoning problems repeatedly across teams to solving them once, centrally. Engineers now fill out a form and receive a pre-configured GitLab repository running FastAPI, complete with LangGraph agent modules, OpenTelemetry tracing, Vault secrets management, service discovery, and built-in evaluation—all from the first commit.
The company describes the shift plainly in its engineering blog: "The reasoning loop took a whole afternoon. The production wrapper took two weeks."
How it works#
LLM-Kit deliberately avoids becoming a rigid platform or domain-specific language. Instead, it wraps Grab's existing infrastructure into reusable scaffolding. Tools aren't hardcoded; agents discover them at runtime from over 50 registered servers using the Model Context Protocol. Model access flows through Grab's OpenAI-compatible gateway, which sits in front of five providers and injects credentials—so application code never needs to know which backend it's using.
Evaluation is also built in, measuring agent responses with ROUGE, BLEU, and a second model acting as a grader.
Why framework, not platform#
Grab chose intentional narrowness. As outlined in the engineering post, a monolithic platform would have "locked teams into rigid assumptions that would soon become outdated. A framework let us meet developers where they already were."
Analyst Kai Waehner has cautioned that agentic AI lock-in runs deeper than API lock-in because it accumulates simultaneously at the model layer, the framework, the runtime, and the patterns teams build. Grab's approach addresses this by keeping the gateway as the only unified layer, while allowing the framework and patterns to remain more loosely coupled.
