Back to Sernia Capital

    SerniaAI operations agent

    One agent across team conversations and scheduled work, with shared company memory and access to the tools we use to run Sernia Capital.

    How the system connects

    Scheduled run

    Proactive checks and follow-ups

    SMS webhook

    Internal team messages on the AI line

    Web chat

    Authenticated team conversations

    Each starts a run with its conversation context

    SerniaAI

    Gathers context, decides what needs doing, and chooses tools. It can act, reply, schedule a follow-up, or leave things alone.

    Read & write

    Shared company memory

    Property knowledge, daily notes, procedures, and reusable skills in a persistent, versioned workspace.

    The same knowledge across people, conversations, and future runs.

    The agent selects tools; application code checks each action

    Agent capabilities

    Computation, research, and context management

    Monty Python sandbox

    Calculations and data transformations through run_python, without filesystem or network access.

    Compaction

    Automatically summarizes large tool results and older conversation history before model requests.

    WebSearch

    Provider-native search, restricted to allowed domains when supported by the selected model.

    WebFetch

    Provider-native page retrieval on supported models. No local fallback when unavailable.

    Business integrations

    Access to company information and operational actions

    Google Drive

    Search and read company files

    Google Calendar

    Read schedules; create or delete events

    Gmail

    Search threads, read and send email

    Quo SMS

    Read conversations and send texts

    These groups describe purpose, not separate permission levels. Code execution and web access are tools too; compaction runs automatically to manage context.

    BashTool is deliberately excluded. The agent is not given a general-purpose Bash tool for host shell commands.

    Permission depends on the action and recipient, not just the tool

    Internal work: act autonomously

    No per-action approval for routine work within the team.

    • Search files, email, calendars, and SMS history
    • Read and update shared memory
    • Send internal emails and texts
    • Manage internal-only calendar events
    • Create and update tasks; coordinate follow-ups
    Execute and return the result to the agent

    External communications: request approval

    The standard path for messages to tenants, vendors, and other external contacts.

    • External SMS and email
    • Calendar changes involving external attendees
    Human-in-the-loop review

    Approve to execute. Reject to stop the proposed action.

    Shared memory carries knowledge between runs; individual conversations retain their own history. Tool results return to the agent so it can continue working. The diagram highlights the main entry points and integrations.

    Where the boundaries sit

    Approval is enforced by the tools, not left to the model’s judgment. Some sensitive internal changes also require review, including task deletion and contact updates or deletion.

    There are narrow external automation paths: fixed, code-owned reminders can send without approval, with recipient eligibility checks and no model-written message text. Leasing email automation also has an operator-controlled approval setting. These are separate from the standard approval path shown above.

    Automated triggers have a kill switch and rate limits. The agent works within authenticated access and tool permissions.