Saltar al contenido
Anuar Harb
ENES

Production agents at Inmercity Labs

Four LLM agent systems in production for clients over WhatsApp, connected to their systems with RAG, tool calling and MCP.

Rol
Co-founder, AI Engineer and Technical Lead
Periodo
2026
Stack
WhatsApp Cloud API, LLM tool calling, RAG, MCP, Programmatic evals
Enlaces
    recurring users on the legal assistant
    300+
    saved per quote by the interior design agent
    ~2 h

    Este caso está disponible solo en inglés.

    The problem

    Our clients needed their customers' questions answered on the channel those customers already use, WhatsApp, with answers that come from the client's own systems instead of from a person copying data between screens.

    Constraints

    • The channel is WhatsApp. No app to install. The agent has to work inside a chat.
    • Answers come from live data. Inventory and calendars change, so the agent reads them at answer time instead of remembering them.
    • Releases can't break a working agent. People use these assistants every day.

    Architecture

    Inmercity Labs agent architecture A user writes on WhatsApp. The WhatsApp Cloud API forwards the message to the agent runtime. The agent loop uses retrieval over client documents, tools for live systems like inventory and calendars, and MCP servers for other client systems. A programmatic eval suite gates every release before deploy. User WhatsApp WhatsApp Cloud API Agent runtime connected to client systems Agent loop LLM with tool calling RAG client docs Tools live data MCP systems for example live inventory and advisors' calendars Programmatic evals every release is gated on the eval suite before it reaches users
    Architecture of the Inmercity Labs agents. Users write on WhatsApp, the Cloud API forwards messages to the agent, which uses retrieval, tools and MCP to reach client systems, and an eval suite gates every release.

    The WhatsApp Cloud API delivers each message to the agent. The agent loop decides between answering, retrieving from the client's documents, calling a tool against a live system, or going through an MCP server connected to another client system.

    The four agents

    Legal assistant. More than 300 recurring users.

    Real estate assistant. Answers from live inventory and books visits on advisors' calendars.

    Interior design quoting agent. Saves about two hours per quote.

    Student agent. Tutoring and enrollment.

    Evaluation

    Every release is gated behind programmatic evals. When a client reports that the agent "feels off," the job is to turn that report into a failing case in the suite, then fix the agent until it passes without breaking the rest.

    Results

    • Four agent systems in production.
    • 300+ recurring users on the legal assistant.
    • About two hours saved per quote on the interior design agent.