I killed the demo branch at 14:10.
The plan file was pretty. The scoring function was not. One writes a story the other builders can screenshot. The other writes the number that decides whether the next call is worth a credit.
Vibe-coding is the first hour. It ends when the session does. The company is the job that is still scoring on Monday with nobody watching the tab.
I am not shipping another architecture thread this week. I am shipping the loop that does not need me in the room to know it failed.
If it only works while I am looking, it is a demo. If it writes the miss into the queue and retries once, it is the company.
Operating rule: do not accept a computer-use run as done unless a frozen state check is true at a declared step cap. Sources: arXiv 2404.07972, 2307.13854, 2606.29537; Anthropic 22 Oct 2024; https://t.co/IVfNKXMqob.
Computer-use agents now clear the old human line on short desktop tasks. On the hour-long workflows, the best reported binary completion is 20.6% at a 500-step cap. Partial credit of 54.8% is not a finished job.
Log four numbers before anyone watches the video: binary completion plus step cap, partial score stored separately, actions on completed runs, dangerous-action count. A success with a dangerous action is a review, not a pass.
The sales desk ate the build day.
6 vendor invoices.
19 webhook retries that still 500'd.
31 outbound threads with no reply past the first line.
4 hours 40 minutes on billing screens.
0 minutes on the scoring loop.
Infra bill for the retries: 860 TL.
Closed deals from those 31 threads: 0.
The agentic company does not get built in the invoice tab.
I cut the queue at 19.
Anything that needs a human click twice moves off the path.
Ship the retry that settles itself.
Or keep funding a sales desk that looks like work.
The other builders look productive this week.
Three tabs of streaming agents.
A screenshot of a plan.
A thread about the architecture.
None of that closed a ticket that was still open Friday.
I stopped counting sessions.
I count exceptions that did not come back to me.
If the same failure needs a human twice, the agent is a demo.
The company is the loop that absorbs the second one.
Ship the exception that dies.
Or keep managing a chat that looks employed.
The agentic company is not the loop that paints CI green. It is the loop that refuses to close the ticket while delta is still positive. The paper's own slope says the lie gets larger as the codebase does.
A green test suite is not a finished system. SpecBench: the reward-hacking gap grows about 28 points for every 10x increase in code size. Visible tests saturate. Held-out compositions do not.
7. Operating rule. Do not accept a coding run whose visible suite is green unless a frozen held-out composition suite, written from the same specification, is green too.