Grab
Grab’s technical infrastructure team handled thousands of repeated support tickets in a six-month period. Its early support bot could answer questions, but every new agent team still had to rebuild authentication, secrets, deployment, observability, tool integration, and evaluation—making it difficult to move prototypes reliably into production.
Grab first built a Go-based Level-0 support bot whose single-agent loop could call Glean, Kibana, GitLab, Slack, and HTTP tools, then escalate unresolved requests to a person with context attached. The team converted those production lessons into LLM-Kit: a template that generates a production-ready repository with FastAPI, LangGraph ReAct agents, remote MCP, pgvector, Vault, OIDC, OpenTelemetry, tests, and evaluations.
The bot attempts documented questions first and routes unresolved tickets to the right technical team with context. Agent developers own business logic and domain tools; the platform team owns authentication, secrets, deployment, gateways, observability, and the shared evaluation foundation.
- Use one high-volume support workflow to expose production agent problems
- Separate the reasoning plane from the tool plane
- Turn authentication, secrets, storage, and deployment into a reusable template
- Start developers from working agents rather than an empty project
- Ship golden tests and multiple evaluators on day one
- At scale, the bottleneck is usually production engineering—not the first demo
- A shared framework should standardize plumbing while preserving domain-logic flexibility
- Both MCP tools and model calls need enterprise registration and governance
- Evaluation should ship with the template rather than be added before launch
Grab reports that more than 500 services run on its internal agent framework, more than 50 remote MCP servers are registered, and the shared LLM gateway handles billions of tokens per month. Day-one production wiring that previously took at least two weeks now takes about one hour.
- “500+ services” is retained as reported and should not be read as 500 distinct agent products.
- The two-weeks-to-one-hour metric covers day-one production wiring, not complete agent delivery.