Grab's LLM-Kit Cuts AI Agent Deployment From Two Weeks to One Hour

Grab has reduced the time it takes to get a new AI agent into production from two weeks to about an hour, using an internal framework called LLM-Kit that now governs more than 500 agent services across the company. The effort, described by engineer Hien Luu, targets a problem familiar to any large engineering org: agents built by different teams with different tools, evaluation methods and secrets handling, which slows every launch and makes reliability hard to reason about. Grab's answer was to stop treating each agent as a bespoke project and start treating it as a workload on shared rails.

LLM-Kit covers three areas that usually cause friction. It standardizes how agents connect to internal services, so a new agent inherits existing integrations instead of rebuilding them; it bakes in evaluation so teams can test behavior before shipping; and it centralizes secret handling, keeping credentials out of individual codebases. The framework also manages the underlying infrastructure, which means teams working on agent logic are not also maintaining deployment plumbing. That division of labor is where most of the two-week-to-one-hour compression comes from.

Two design choices stand out. The first is runtime tool discovery: rather than hardcoding a fixed list of tools when an agent is built, the agent can find available tools while it runs, so capabilities added later are reachable without a redeploy. The second is flexible model integration, letting teams swap or combine models without rewriting the surrounding agent. Together they push configuration decisions out of the build step, which is precisely where slow release cycles tend to form.

The scale matters as much as the speed. Five hundred services is past the point where a framework is a side project, and past the point where a single team can enforce standards by hand. Grab kept operational control centralized while giving product teams room to move, a balance that enterprise agent platforms often struggle to strike. For companies running dozens of pilots that never reach production, the interesting claim is not the hour, it is that standardization and speed were not treated as opposites.

What Grab has not published is the harder half of the story: how LLM-Kit handles evaluation quality, failure modes and cost when agents are generated that quickly. A one-hour deployment pipeline is only valuable if the agents it produces are worth deploying, and evaluation is the usual bottleneck. The open question for the rest of the industry is whether this pattern generalizes outside a company with Grab's engineering depth, or whether it depends on infrastructure most organizations have yet to build.