Grok Galaxy day one: how to direct and supervise bots

Peter Dedene @dedeneOriginally posted 2026.09.16

Open original on X

I sat through Grok Galaxy Day 1, close to nine hours. 🤯 At one point, the team noticed they already had 13 PRs. Some were prototypes they'd forgotten asking a bot to make. Hours later, they were still deciding who they were building for. That was the part I found most useful: watching how people direct the work, keep the bots on track, and check what came back. These are the things that stuck with me. 👇 Specialist bots are how you keep context from turning into mush. In the engineering session, the presenter split UI, infra, and DevX even though they share a model and can cover each other's tasks. Each bot has a different pipeline and a different pile of corrections. I would pick a bot's scope by what it should remember on the next job. Someone still has to supervise the coding agents. The engineering presenter described juggling about 15 cloud agents: check progress, write follow-ups, interrupt work that had drifted. Their Grok Bot reads a worker's transcript and asks for missing screenshots or perf numbers before it brings anything back. The handoff includes definition of done. "Hurry up" made agents guess and skip steps. Their P0 is: look in every five minutes, intervene when a worker is stuck or wasting time. An ops bot writes that into a shared playbook and tells the other engineers. Verification needs its own tooling. The verification skill they described combines a CLI for driving the app with a feature map of what exists and how to reach it. Without that, each agent invents a test script or asks you to click around. They still had to go back when PRs lacked the screenshots and videos they wanted. Writing the skill and checking that agents use it are two jobs. The data bot can reject the task. In the PM demo they misread a funnel and asked for seat-selection work. The data bot said the leak was "search to fare selection", then passed the corrected analysis to the product bot. The useful collaboration happened before anyone built the wrong fix. Persistent memory needs a feedback loop. A customer bot reviews what happened on calls. A writing bot compares its draft with the message that went out. Those edits are preferences you will never put in a style guide. A manager bot can still remember deleted teammates, so the instructions have to be updated. A routine every fifteen minutes is about a hundred runs a day. Telling it to stay quiet does not cancel those runs. Audit your schedules or you're wasting tokens. Prefer an incoming signal. Use an API for repeated work. The browser path costs time and tokens. You do not have to copy someone else's org chart. One speaker talks to specialists. Another uses a chief of staff over a larger roster. Both showed up on the same day. Stay as involved as you want, then change it when coordination costs more than it saves. Seeing how people work with their bots, was well worth the time. More live builds, more messy demos, more of this please! 🔥

This post quotes another post: x.com/poteto/status/2099875666313527436