Grab unifies over 500 agent services on LLM-Kit

Grab's write-up centers on a split that is easy to miss. The inference loop itself, meaning the core logic where an AI reasons, calls tools, and answers, took a single afternoon to write. The surrounding production shell, covering where credentials live, how logs and traces are wired, service discovery registration, and evaluation, took two weeks. Grab's move was to stop letting each team decide those elements on its own. A form now returns a GitLab repository with secret management and service discovery already wired, plus an evaluation endpoint attached from the first commit. Notably, Grab did not invent a new abstraction layer or a dedicated language; it built scaffolding once around infrastructure the company already had.

What moved to the center is the calling interface. Model calls pass through one OpenAI-compatible gateway that bundles five providers and inserts credentials, so app code does not hard-code a provider name. Tools are not bundled in advance; they are fetched at runtime from over 50 MCP servers, a shared connection standard for letting AI use tools. Adding a new model or tool needs no redeploy, only a registration or config change. The layers above that, however, stayed with the teams. Five hundred services can share the same scaffolding, yet prompt quality and the criteria for what counts as improvement still exist once per team.

The article also notes a stated tension. Grab said it chose a framework over a platform to avoid locking teams in, then later wrote that at 500 services the problem had shifted from framework to platform. The write-up reads this sequence as deliberate: calling it a framework got 500 services onto it, and that scale then justified centralizing the shared pieces. The piece also cites a separate analyst's warning that agent lock-in is harder to unwind than API lock-in because it accumulates across four layers at once. For a company weighing whether to build the same thing, the article argues the first move is a single model calling interface, which alone covers cost visibility, provider swapping, and credential aggregation, while MCP registries and evaluation infrastructure can wait until enough services share them.