Agent Realtor

Problem

Real estate agents juggle dozens of contacts, follow-ups, and scheduling tasks across email, text, and social media. They lose deals to missed follow-ups and forgotten context. Existing CRMs are expensive, clunky, and require manual data entry that solo agents won't do consistently.

Approach

Built an autonomous operations agent that plugs into the tools realtors already use, with no new apps to learn. The system observes Gmail via Pub/Sub push notifications, with a prefilter service that blocks promotional noise before the agent even wakes. A 3-tier memory system gives the agent persistent context tailored to real estate: a curated set of facts loaded into every prompt, an archive of observations searchable via semantic embeddings, and a structured contact graph with full interaction history. The agent graduates its own knowledge between tiers, keeping a pressure valve that forces it to curate rather than accumulate. Calendar events are cross-referenced with known contacts for morning briefings. A 4-layer cost optimization stack (pre-check skipping, light context for quiet cycles, daily session resets, cheaper models for routine work) compounds to roughly 99% savings on idle heartbeats. Each user runs in an isolated rootless Podman container with their own SQLite database, OAuth credentials, and Telegram bot. The infrastructure layer is being extracted into HiveClaw, a product-agnostic Go runtime targeting sub-50MB memory per instance, designed to scale to 50+ persistent agents on a single 8GB host.

Result

Per-cycle cost dropped from $0.10 to under $0.01 after optimization. A working realtor processing ~250 emails/month runs at roughly $2/month in extraction costs. The agent handles daily operational overhead autonomously: contact tracking, follow-up reminders, calendar awareness, morning briefings, with zero manual CRM entry. The infrastructure layer is being generalized into a framework that replaces 1.5GB-per-container runtimes with a sub-50MB Go process.

Stack

Links