
Feb 23, 2026 · 8 min
How Shadow Inbox was built with OpenClaw: an AI agent factory case study
We built Shadow Inbox with OpenClaw, an AI agent factory. Here's the agent role split, the 1-week MVP, the surprises, and what we'd do differently.
arthur

We built Shadow Inbox with OpenClaw, an AI agent factory. Here's the agent role split, the 1-week MVP, the surprises, and what we'd do differently.

AI reply generators face a dilemma: fast and cheap is templated-with-extra-steps, slow and personalized is barely better. The third mode no one is building.

Classifying buying intent with LLMs is mostly about evidence quotation, temperature pinning, and not asking the model dumb questions. Here's what worked for us.