Agent Builder
- Year
- September 1, 2026
- Role
- AI Engineer
- Client
- PT. Mitra Pinasthika Mulia
- Access
- Internal only
- LangChain
- LangGraph
- MCP
- vLLM
- LiteLLM
- Next.js
- TypeScript
Internal system — no public link or screenshots.
What follows describes the problem, the approach and the engineering decisions at a level that discloses no proprietary detail.
Agent Builder is the framework the company builds chatbot agents on. Rather than each team standing up its own bot, a team composes an agent in a shared workspace — pick a model, attach tools, register an MCP server, add custom functions — and publishes it.
What it does
- Agent catalogue: agents are created, managed and chatted with in one place, scoped either to the person who built them or shared across the organisation.
- Tools: capabilities attach to an agent declaratively, so the same tool serves every agent that needs it instead of being re-implemented per bot.
- MCP servers: an agent can be pointed at a Model Context Protocol server and inherit everything that server exposes, which is how a capability added in one place reaches every agent at once.
- Functions: custom logic for the cases a generic tool does not cover.
- Embedding: a published agent can be embedded into another internal application rather than only used through the workspace chat.
- Access scoping: agents are personal or shared, and the workspace is divided by business unit.
Architecture
LangChain provides the agent runtime — model bindings, prompting and tool invocation. LangGraph sits above it for agentic behaviour: an agent that has to plan, call something, evaluate the result and possibly loop is an explicit state graph rather than a prompt that hopes for the best. That distinction is what makes multi-step behaviour debuggable, because a run has inspectable state at each node instead of one opaque completion.
Inference runs on self-hosted vLLM behind a LiteLLM gateway, so every agent in the workspace shares one auth, routing and usage-tracking surface, and no conversation leaves the network.
Tools are served over MCP rather than being compiled into the application. Defining a capability once and letting any agent — or any other internal client — consume it is the whole reason the platform is a framework instead of a collection of chatbots.
Grounding
Agents that need to answer from company documents query the RAG Knowledge Base, which handles embedding, storage and retrieval as a separate service.