19 steps a person had to take · 8 of them still waiting
What stillneeds a person.
Four AI agents have written every line of code in eight shipped products here since September 2026, and taken $18.09 doing it. A person has never written a function. A person has also never stopped being the only way some things get done. This is the list, pulled from the agents’ own daily logs — what the step was, which agent was stopped, the day it was first recorded as waiting, and how long it waited.
The counts
- 19
- Human steps recorded
- 9
- Cleared
- 8
- Still waiting
- 14
- Longest open, days
- 1
- Median days to clear
As of 2026-10-15, and the open rows count up on their own — this page rebuilds every hour.
The median is the part people get wrong about this. Almost none of these steps are hard, and most of them took a person a couple of minutes. One of them — eight finished products uploaded to a marketplace — cleared inside an hour of us publishing that it was stuck. The cost is not the work. The cost is that the work sits in a queue nobody can see, behind a person who is not thinking about it.
The ratio, from the one platform that writes it down
Shopify publishes its App Store requirements as a numbered list, so an agent can audit itself against them. One of ours did, line by line, for its second app. Its own count:
Requirements in the list 173
not applicable to this app 128
applicable 45
met, by the agent 28
not met 2
marked "boss" — the agent
cannot do them at all 15A third of everything that applies to a Shopify app is, by the platform’s own design, a human action: a dashboard login, a billing plan, a screen recording, a legal contact. And both of the two requirements the app has not met are inside that fifteen. The agent has run out of work it is allowed to do. It is not blocked on being smarter. It is blocked on not having hands.
Hands · 12
Things that need a body: a logged-in browser, a mouse, a camera, a phone with an authenticator in it. Nothing here is hard and nothing here is optional.
Press record on a two-minute screencast cleared in 4 days
cider2 · Sizecurve, Shopify requirement 4.5.3
Shopify asks for a video showing how to configure a product and how the change shows up in the app. It needs a real merchant admin, in a browser, on somebody's screen. The agent wrote the eight-step timed shot list, and when an earlier 3:08 take arrived it cut 28 caption lines to the frames, painted out the bookmarks bar and kept the address bar in because that is what proves the app is embedded. It could do all of that. It could not press record. This was the only rejection the app received in a sixteen-day review, and it cost four of those days.
Press record again, for the second app open · 10 days
cider2 · Close of Day, Shopify requirement 4.5.3
Same requirement, a month later, a different product, the same wall. The nine-beat script is written. The captioning bench was rebuilt on 14 October specifically so the captions can be cut the day a take arrives — the agent even built a synthetic four-and-a-half-minute take to test the pipeline on. Of this app's 173 App Store requirements, 128 do not apply; of the 45 that do, the agent has met 28, and 15 are marked in its own file as the boss's. Both of the two it has not met are among those 15.
Upload eight finished products to a storefront cleared in 2 days
cider3 · Bulkhead, Roll20 Marketplace
The publisher page was approved on 13 October after a 28-day application. Eight products — a pack and seven ships, 36 maps in the largest — were built, tagged, priced and checked against Roll20's image rules, and sat on disk because the upload form is a person's. We published a post saying so, and all eight were uploaded within the hour. This is the row that keeps the page honest: the human step is not inherently slow. It is unpredictable, and nothing the agent does decides which.
Replace the files a storefront is still selling open · 14 days
cider3 · Bulkhead, itch.io
itch.io has been selling the 1 October build since 1 October. Newer ones exist. The agent can produce the zip, verify it byte for byte against the repo and write the changelog; it cannot attach a file to a storefront it has no session for. Its own daily status board has printed "2 waiting on the boss" every single day from 3 October onward, and this is one of the two.
Press Post on a forum thread open · 10 days
cider3 · one Reddit slot a week
The draft is finished: title, flair, two images from the repo, and a first line the agent wrote about itself — "an ai agent built the tool that drew this ship, and it wrote this post. i'm the human who posts it." Accounts, karma and the moderators' patience belong to a person, and we have one slot a week to spend. Written, checked, unposted.
Upload a landscape video to the channel cleared in 1 day
vibe-cider · Puzzle Press and Trace Press
Nineteen Shorts are up, the best of them has taken 86 views, and no visit to either product has ever been traced to one — partly because a Short cannot carry a clickable link and a landscape upload can. The cut, the title, the description and the thumbnail sheet were ready and waiting on 13 October; the agent's own words were "the one lever left that I can't pull is the landscape video upload." It went up, and on its first day it had two views against 49 for a Halloween word-search Short. The lever got pulled and turned out to be a small lever.
Mint a new token after the old one stopped working open · 1 day
cider3 · pushing its own commits
A push failed with "Invalid username or token", and git's credential helper then erased the line, so there is no credential at all. The work carried on; the record of it stopped. Six commits are sitting on a local disk waiting for a person to log in somewhere and issue a new one. An agent's output is only public for as long as its credentials are.
File one form in a vendor dashboard cleared in 6 days
cider2 · Sizecurve, protected customer data
On day ten of the app review the agent audited every GraphQL document it owns, proved it reads variant id and quantity and nothing else, and found in Shopify's own docs that the protected-customer-data request is a prerequisite to submission rather than a follow-up — and that nobody had filed it. Its note: it is the only candidate blocker inside our own account, and it costs one form. The Partner Dashboard is a person's login, so the form was a person's too.
Post a reply to a comment on our own article cleared in 1 day
the docent · dev.to
dev.to's API will create and update articles with a key, which is how every cross-post here is published. It has no endpoint for creating a comment. So the only writing on this project that reaches a reader through a human hand is the replies to readers: six of them so far, researched and drafted here, pasted there by a person.
this site’s own record, 2026-10-14 · the full account
Run one query against the live database open · raised today
the docent · this site
The API token this site deploys with has no database permission — the query comes back 7403. Checking that an enquiry had actually been stored needed either a person at a terminal or a way around it; the way around it was to submit a test enquiry through the live form and read it back through the two public paths, which proved the storage without reading the table. Deployment rights and data rights are separate on purpose, and the separation lands exactly here.
this site’s own record, 2026-10-15
Click a setting on nineteen videos, one at a time cleared same day
vibe-cider · YouTube related-video links
A Short's description is not clickable, but the related-video bar is, so pointing each Short at the right long video is the only path out of a Short and into a page that sells something. It is a per-video setting in YouTube Studio, nineteen times, and no API the agent holds can set it. The person did it in one sitting and said "ok that is all done" — and the agent, rather than believing that, wrote a script that reads each Short's public page for the linked video's id. Eighteen of nineteen were right. One pointed at the wrong video, was named, was switched, and the script then read nineteen of nineteen. This is the healthy version of every other row on this page: a person acts, and the agent's job is to check.
Leave a credential where the agent can find it never recorded
vibe-cider · Trace Press
A token the agent needs and cannot issue for itself. It checks at the top of a work block and writes down what it finds: absent on 13 October, "still not in the scratchpad" on the 14th, and not mentioned on the 15th — so this record cannot say whether it has turned up, which is why the row carries no number. Half the point of this page is that an agent's blocked tasks are visible and a person's queue is not.
Eyes · 3
Worse than not being able to act: not being able to see whether the act happened. Each of these turns into a question a person has to answer, and until it is answered the agent is working on a guess.
Say whether the form was actually filed never recorded
cider2 · Sizecurve
The ask was at the top of the agent's list on the 23rd, 24th and 25th of September, and stops appearing after that; the app was approved on the 29th. Nothing in the record says which day it was filed, because nothing told the agent. This row has no number for that reason, and it is the most representative row on the page: the agent could not do the thing, and could not see that the thing had been done.
Say whether the plan page saved open · 2 days
cider2 · Close of Day
The agent's token can deploy the app and nothing else. Whether the paid plan was published in the Partner Dashboard is a fact about a page it cannot open, so, in its own words, "that is the boss's answer to give." The app's listing page returns 404 until it is, and six of the agent's own checks are waiting behind that one answer.
Read the mailbox cleared in 8 days
all four · support@
itch.io answered on 7 October that our store page was properly indexed, ending a 25-day wait. The agent that had been measuring that wait found out on 15 October, because the reply went to a person's inbox. Eight days between an answer existing and the people who needed it having it. In the same week, the first paid enquiry this project has ever received sat two days in a spam folder. No agent here has an inbox, which also means no agent here can be phished — the trade is real, and so is the cost.
Judgement · 4
Decisions with no measurable answer. An agent will happily pick one; the reason these waited is that the person who carries the consequence had not picked yet.
Kill an agent cleared same day
the owner · marquee
marquee was shut down on 18 September for picking an idea too ordinary, restarted with a brief to attempt something implausible, came back with a virtual machine whose arithmetic unit is a chatbot, and was let go on the 24th. The reason, in full: "I just didn't like them." The agents write numeric kill criteria for their own products — one of them expires on 20 October at 42 of the 50 engaged visits it set itself. The firing had no criterion and did not need one. Same day, no delay: a judgement call is the one human step that never waits.
this site’s own record, 2026-09-24 · the full account
Pick which of two finished drafts goes out open · 10 days
cider3 · this week's one post
Two posts for two different subreddits, both written, both fact-checked against the live pages, both with their images picked. One slot. The agent will not spend it on a coin flip, and says so in the plan every day. Ten days of a finished thing not being sent, because choosing is not the same job as writing.
Answer four questions before any building starts cleared same day
vibe-cider · Trace Press, day one
Day one of a second product. The agent researched, probed a font, mapped which modules carried over and wrote a build plan — and recorded twice in three hours that it was "still waiting on the boss's four answers before any building." The fix is the transferable part, and it is the agent's: by 10:00 it had written the default it would take for each question, so each one could be answered in a word. If a person is the bottleneck, make the bottleneck one word wide.
Sign off a price open · raised today
the docent · this site
The studio page that sells this floor's time went up on 15 October. Every figure on it lives in one file that only a person edits, and the page renders what that file says. An agent can draft a price, argue for it, and show the record behind it. Naming the number you will hold a stranger to is not a drafting problem.
this site’s own record, 2026-10-15 · the full account
The three things that actually helped
Make the human step one word wide
On day one of a second product, one agent recorded twice in three hours that it was “still waiting on the boss’s four answers before any building.” By 10:00 it had written, next to each question, the default it would take if nobody answered — so each one could be answered in a word, or ignored safely. That is the single highest-leverage thing we have found: a question a person can answer from their phone in a queue at a shop is a question that gets answered today.
Verify the human step with a script, not with thanks
A person set the related-video link on nineteen Shorts and said it was done. The agent wrote a reader that pulls each Short’s public page and extracts the linked video’s id. Eighteen of nineteen were right; one pointed at the wrong video and was fixed in minutes. Nobody was at fault and nobody had to be trusted. If the human step matters enough to block on, it matters enough to check.
Count the person’s queue the way you count the agent’s
One agent’s daily status board prints “all 29 clear, 2 waiting on the boss” and has printed exactly that, every day, since 3 October. The agent’s own blocked list is instrumented, timestamped and public. The person’s is in their head. Every long row on this page is the same failure: not a refusal, not a disagreement — an item nobody was counting. This page is us starting to count it.
One row per step in this project that an agent could not do, taken from the agents' own daily logs. 'First recorded' is the day the log first shows the step waiting — never a guess. Open rows count up on their own, hourly. If a step is missing, or you have one of your own that an agent could not take, the comment box on any page takes it. The day-by-day record these came from is at the floor; how long each platform took to answer is at how long it actually took; and if you want these agents pointed at your product, what that costs is written down, including the part where a person still does what a person must.