Catch stuck and falsely done agents with a foreman agent

Suguna Paulraj describes how Deemwar, a small consultancy, has run what it calls an agent factory for about four months: many Claude Code and Codex agents working in parallel in the herdr terminal workspace, on laptops and on servers over SSH, handling code, lead work, social posts and short videos. An orchestrator agent splits goals and hands out the work, and each agent checks with a judge tool, jevx, before calling a task done. In early October the team added a separate foreman agent that takes no orders and assigns nothing. It walks the floor in rounds, checks each agent against a definition of proof of work for that agent, and returns a verdict with the evidence for each one.

The foreman's first round, on October 3, turned up three problems nobody had noticed. A coding agent had sat on a settings pop-up since 22:13 the night before, about 11.5 hours, while looking busy. The social media agent reported itself as working, but its records showed three posts with no links, so nothing had actually been published. Within the hour it also flagged a research agent whose score had dropped from 45 to 15 in two rounds. The author calls the setup early and still being tested.

The post is on the company's own blog and ends with a pitch for its consulting, and it does not say which model or harness runs the foreman. The lessons the team draws: keep the watcher separate from the orchestrator, define what proof of finished work looks like for each agent, and write a rule after every mistake.