Bobby.

An agent built around daily operations

Bobby supports a property team with the repeated work surrounding lettings: finding potential leads, keeping up with enquiries, preparing messages, coordinating viewings, and tracking the next action on a case.

I extended the Hermes agent framework with Python tools, persistent records, and scheduled workflows. Bobby runs on an Orgo cloud computer, uses a browser to work with SpareRoom, and connects language models through OpenRouter. The team can interact through Telegram and Bobby CRM.

From conversations to workflows

• Outreach across listings. Scheduled searches collect candidates for active listings, track their review state, and maintain a delivery queue.
• An actionable inbox. Conversation checks identify messages needing a reply and summarize what the person is asking, alongside links back to the conversation.
• Viewing coordination. Stored viewing slots support offers, confirmations, and follow-ups, with outcomes reflected in the lead’s record.
• Shared operational memory. SQLite-backed records retain leads, listings, cases, approvals, and activity across scheduled runs and team conversations.

Making automation controllable

Bobby’s workflows combine model-generated interpretation with explicit application rules. Approval records bind a proposed message to its conversation and text. Delivery checks account for sending windows, pacing, duplicate messages, and whether the conversation has changed since the draft was prepared.

The system also handles interrupted workflows: locks prevent overlapping jobs, failed chat checks can be retried, and maintenance or session problems can pause work for later recovery. Human review and handoff remain part of the workflow where judgment or intervention is needed.

Connecting the agent to the team

The CRM integration publishes changed records and receives commands through outbound polling. Team actions use Bobby’s existing functions and validation, with results returned to the workspace. Notion now serves as a read-only mirror of that operational state.

The most substantial work has been around the model: maintaining state, coordinating background jobs, checking the browser’s current state before acting, and giving the team a clear way to review or take over a workflow.