GitHubWhat they did
Alice Moore, who works on DevRel at Builder.io, set up 16 GrokBot bots, one per area of her work (Inbox, GitHub, Bugs, Blog, Research and so on), plus a Config bot that set up the others and a Vault bot that alone could write to her notes. Her coding ran in T3 Code, a desktop app that runs Claude Code and Codex threads side by side. Rather than expose T3 Code through a public tunnel, a small plain script on her PC called a GrokBot webhook whenever a coding thread finished, failed or needed her, which woke a bot that could report back or act on the PC through the GrokBot desktop app. A GitHub bot watched her PRs and, once a review was approved, told the owning coding thread to land it. Three times each weekday the inbox, calendar and journal bots filed reports that Vault turned into a ranked priorities board kept in a documents app.
How it went
In one week she merged 78 pull requests, against a previous best of 35, and one PR landed two minutes after approval. But follow-ups slipped: on one Friday 24 coding threads finished, she heard about 17 on time, 3 late and 4 never, and a job failed twice with no alert. A merge hold she asked for never reached the threads, and a PR merged past a failing check 46 minutes later. About 15% of group-chat replies went to the wrong channel. A filter meant to cut event traffic by 80% was built but never switched on, so bots woke 500 to 750 times a day, and over a weekend they spent $100 and hit the spend limit, on the $200 plan. Scheduled jobs ran 12 to 80 minutes late. She stopped using it.
Worth knowing
Every wake is a paid model turn, even when the bot decides it has nothing to say, so event volume drives cost. Her fix is to let plain code route events and track follow-ups, and only wake a model when there is real work. The post is on the blog of Builder.io, which sells Agent-Native, the documents app she used for the priorities board, and she says she is building an open-source alternative.
Discord