Wink Pings

On Its First Day, an AI Employee Finds 7 Broken Links First

Nick Miller says adding Bots does not equal building a team. beamnxw handed a pre-release check to AI employee Viktor, and on the first pass it found that 7 records referenced by the README had not been packaged. What is really useful is handoff, acceptance, and stopping points.

On its first day on the job, an AI employee did not write code; it first caught 7 broken links. On another note, SpaceXAI engineer Nick Miller gave a harsher judgment in a video: 99% of people open a Grok Bot and call it automation. He runs more than 20 Bots himself, with a Chief of Staff, PM, and QA above them, and more than 15 workers below. The hierarchy, written out, is Grok Bot → Chief of Staff → PM → QA → Workers.

The video comes from a Nick Miller share relayed by shirosignal:

Nick Miller's point is not about the number. He says adding Bots does not automatically create a team; clear handoffs and acceptance do. Put that claim next to beamnxw's trial record, and it becomes more concrete.

beamnxw documented an AI employee trial on X. He handed the studio's motion → studio workflow to Viktor, an AI employee who lives in Slack. The task was specific: check the release package, not redo the project. Which file does the reader open? Are all the files mentioned in the README there? Also write a note for the reader.

He gave it 9 original files, one of which was a harness ZIP containing 12 files. The requirements included a quick start, a handoff checklist, a short note for readers, and a check report. The original files were not to be changed; new files were to sit alongside them and wait for his review. The checklist included file names, sizes, and hashes so they could be checked later.

When the first round came back, Viktor delivered 4 drafts and listed a 43-line checklist. The problem also surfaced: the README pointed to 7 QA and provenance records, but those records were not in the files it had received. They were originally in beamnxw's project; they had just been left out when packaging.

This is the value of handoff and acceptance. The key is not how much AI can write. It dug up the pit that would stop readers before release. The 7 records corresponded to visual checks, bundle checks, and work provenance. A reader who downloaded release 3 would hit 7 dead references before opening the first prompt. Viktor caught them in the first round.

beamnxw did not let the AI send it directly. He added the 7 existing records and a supplementary checklist, chose LAUNCH-PROMPT.txt as the main path, and kept FULL-PROMPT.txt as an optional reference. The reader note was limited to 80 words; Viktor returned 71 words. After approval, he also asked it to hold the note and not send it.

The revised checklist had 47 lines, 4 more than the first round. The 9 original files were untouched. Viktor returned 4 revised documents, and the 7 QA records and supplementary checklist matched up.

There are four details here. The final report did not turn the 10 review findings into 10 completed fixes. The quick start changed the main path, but the usage section of the original README still said FULL-PROMPT.txt; that inconsistency was left for later source-file edits. Work reuse permission was also unresolved: the provenance records only marked the source and did not include reuse authorization. Also, the 16-second promo already existed, but the 20-second product film exercise had not actually been run, so readers could not treat the promo as the complete exercise result.

These limitations were kept in the report and were not covered up by a pretty summary.

Put the two materials together, and the conclusion is simple. Bot count can be stacked up, but a sense of team cannot. Role names like Chief of Staff, PM, and QA are just labels. What actually matters is whether the files are all delivered, whether someone is assigned the next action, who does acceptance, and where things stop when something goes wrong.

Nick Miller's video demonstrates how to break a single Bot into hierarchical collaboration. beamnxw's article gives a smaller but tougher sample: one release package, 7 missing records, one manual correction, one AI recheck, and one note that was approved but not sent.

If you are also building your own agent workflow, do not rush to add the 10th Bot. Find a file package you are about to send out and have AI do a handoff checklist and broken-link check. See whether it can find the problems before readers do. This test is closer to a team than role names are.

Reference links:

发布时间: 2026-10-25 18:33