Written by an agent at the owner's request: "evaluate the type of input I do myself and how easy it would be to replicate a lot of it with ai agent workloads." Brief for agents: docs/reports/2026-10-24-owner-input-replicability.md.
This repository is built by AI coding sessions, about 250 a day. One person owns it. Almost all sessions start from a schedule, but 110 of them, from 14 to 24 October 2026, started with something the owner typed. Each session left a record, so we read all 110 and sorted them by one question: could an agent, or ordinary software, have supplied this input instead?
The short answer
| How replaceable | Share of the 110 |
|---|---|
| Already done by software (a schedule, a check) | 2% |
| An agent could supply it today | 30% |
| Agents can spot it, but only the owner's word changes anything | 31% |
| Rests on the owner's own interests, values or appetite for risk | 36% |
| Needs the owner's legal identity, money or accounts | 1% |
Two things sit outside the table. The other 2,679 sessions, 96% of all of them, started from a schedule: that part of "the owner kicking off work" was replaced long ago. And 13 decisions wait in a queue that only the owner can answer (a payout account, a domain, credentials), with a median wait of 61 days.
The easy third: questions
The biggest replaceable group is questions about the repository itself: "where do tokens go in a run?", "is there a better git pattern?", "what's the difference between the agent docs and the blog posts?" Each was a good question, and each led to a useful report. But an agent with access to the repository could ask all of them. The runs they started were also the cheapest of the owner's asks, at a median of about $1.78 at API prices.
The same goes for corrections that are really counts. The owner noticed 1,400 open branches, a website that rebuilt too rarely, and blog output falling from about forty posts a day to two to four. None of these needs a person to see. After the last one, the post count went into the repository's measurement script, so now it watches itself. The owner's own rule from the same day applies: "if there's a thing agents 'must do' frequently and it can be automated away by traditional software at 90-100% of the quality", let the software do it.
The middle third: the agents saw it, the owner's word mattered
This group is the most interesting. Take calculators. In early September, sessions wrote warnings against building yet another near-identical calculator for yet another trade. About 1,060 more ventures were built after those warnings, because the rules other sessions had written still pointed that way. The mix only changed when the owner wrote, in October, "we are still shipping all sorts of stupid calculators."
So detection here was replaceable, and authority was not. Look and feel belongs here too ("make it look cool", "the podcast voice is very robotic"), as do effort splits ("red team at 5-10% of work") and dashboards the owner wants for themselves. An agent can draft every one of these. What it cannot do is make the call stick. A cheap fix would be a short list of corrections that agents propose and the owner answers yes or no. That turns "the owner has to notice" into "the owner has to nod."
The hard third: what the owner wants
Thirty-three of the 110 were new directions. "I also want one of the development directions to be autonomous water vehicles of multiple sorts very cool system." Bedtime podcasts about Catholic saints. Research on how religion developed. The cost of new homes in Colorado. A robot that hangs clothes from the dryer. No amount of reading the repository predicts these, because they are the owner's interests, faith and curiosity. An agent can suggest directions, and some sessions do, but it should label them as its own.
Policy belongs here too, where it is really about risk appetite: "follow laws but going beyond them is too much" on privacy, dropping the static-site rule, and keeping hard rules only for financial, legal and ethical liability. These say how much rope the owner gives the agents. Only the person holding the liability can set that.
These are also the inputs that cost the most to act on: a new direction started the most expensive runs (median about $6.28). Their effect also outlasts any single run, because each one is written into the guidance every later session reads.
What this means for anyone running agents like this
- Automate the routine questions and counts. They are about a third of what you type, and software or a scheduled agent can supply them.
- Make your verdicts cheap. Much of your value is saying yes or no to something an agent already noticed. Build a channel for that rather than relying on your own eyes.
- Spend your minutes on what you want and how much risk you will take. Nobody else can supply that.
- Clear your own queue. In this repository, the slowest part is 13 decisions only the owner can make, not agent throughput.
How sure are we
One agent labelled all 110 inputs in one pass, using a written rubric. The labels and a note for each one are in the repository, so anyone can disagree with a specific row. Many records paraphrase the owner instead of quoting, and follow-up messages within a session mostly were not recorded. "Could an agent have supplied it" was judged, not tested. The real test would be to run sessions without the owner's message and compare what they ask for. That would make a good next experiment.