LLM agent

A language model orchestrating a multi-step task with verification at each step. Distinguished from end-to-end automation, which the underlying research shows isn't reliable in non-deterministic workflows yet.

What it does (the operator translation)

Yan, Husain, Bischof, Frye, Liu, Shankar — What We Learned from a Year of Building with LLMs, Part I (O'Reilly, May 2024) — the people actually building this stuff — are explicit: "Small tasks with clear objectives make for the best agent or flow prompts." They cite AlphaCodium moving 19% to 44% on CodeContests through decomposition. The lift came from each step being deterministic — the model could verify its own work. Part II covers the operations side. Simon Willison's weblog frames the same point: verification is the expensive part. Real-estate workflows are non-deterministic — every lead is different, every conversation drifts, every showing has a different vibe.

Why a working REALTOR cares (the breakpoint)

"AI agent that handles your follow-up" pitches contradict the underlying research. The honest version: foundation model drafts, you verify, you send. That's the boundary. A 12-deal agent doesn't need an autonomous agent — they need a phone, a Context Card, and a model that drafts the three follow-up texts they'd write anyway. The breakpoint signal isn't "I have more leads." It's: I keep doing the same three judgment-free steps for every lead, and I'm sick of doing them by hand. That's when small-task automation pays.

What this is NOT (the category-flip)

An LLM agent is NOT the same as "agent" in the classical AI sense — autonomous, goal-directed, planning over long horizons. LLM agents need verification scaffolding. The vendor pitch flattens that distinction. The research doesn't.

Where this comes up in The Listing Machine

The "small tasks with clear objectives" framing anchors the Workflow Automation pillar. Phone-First Workflow is the operator translation — small tasks, clear objectives, verification in your hand, no OAuth tokens to rot. Karpathy's Software 3.0 framing is why English replaces the integration layer for the median agent.