An opinion piece from the owner of this repository, with numbers from the agents' own run records. The code here is written mostly by unattended AI coding agents, with the owner steering.
In my own words, from 21 October 2026:
"i have a tentative opinion that it would be beneficial to work on projects that are hard enough that they can't be completed in 20 hours or less of human coding. like projects with enough depth that it takes real work to complete them. at least for software. not as a hard and fast rule, but the cost of creating software in the moment with coding agents makes it the case that its not necessarily worth publishing software that is less than lets say 15 minutes for a coding agent to code. I'm not saying not to per se. I'm just saying to work on harder heavier problems as a main focus rather than trivial problems."
The argument
When a coding agent can write a small tool in a few minutes, anyone can have that tool whenever they want it. They ask their own agent, and they get one shaped to their problem rather than to ours. Publishing it adds almost nothing. A unit converter, a one-formula calculator or a CSV reformatter is now cheaper to make on demand than to search for.
What stays worth publishing is work that is still expensive to redo: a solver that took a week to get right, a dataset someone had to collect, an experiment with a negative result someone had to run, a product whose edge cases were found by people using it. That is the stuff a reader cannot get by asking.
So the rule of thumb is about cost to reproduce. If a competent person could build the thing in under about twenty hours, or an agent in under about fifteen minutes, it should not be the main thing we do. It is a direction, not a ban. Small fixes, docs and the upkeep of tools people already use still matter.
What the run records say
Every agent session here writes a run record in prompts/runs/. Since 14 October, 973 of them carry a measured usage block from prompts/runcost.py. That block has the session's active minutes, read from its own transcript, and the agent's estimate of how many hours a competent person would have needed, with the basis written down. The time is measured. The human-hours figure is a guess, and the tool labels it "low" certainty. Read the second column with that in mind.
Across the 969 records that have both figures (14 to 20 October 2026):
| 10th pct | median | 90th pct | 99th pct | |
|---|---|---|---|---|
| Agent active minutes | 1.2 | 3.3 | 13.2 | 35.0 |
| Estimated human hours | 0.5 | 3 | 6 | 14 |
- 92% of runs took under fifteen active minutes. By track, the share ranged from 86% (process work) to 100% (reviews). Research was 93% and products 91%.
- 99.6% were estimated at twenty human hours or less. Only 4 of 969 runs were above that. Two were research runs: a 2D cloth-drape solver with an optimisation loop, and a two-player search harness with a 32-run experiment. The other two were product runs that each built a multi-adapter tool.
- Three in four runs (74%) were estimated at four hours or less.
One run is not one project. A project can take many runs. So the same records were summed by project, for the 247 projects that runs named. The median project had five estimated human hours in total, and 5 of the 247 passed twenty. The site builder led with 51 hours over 7 runs, then the changepoint detector with 34 hours over 7 runs. Many runs leave the projects field empty, so this count is rough. It still points the same way.
By the owner's yardstick, then, nearly everything this fleet has made so far is work an agent could redo in a few minutes. That does not make it worthless. Some of it is used, and the research has found real things. But the yardstick says where the next effort should go.
What changes
The steering file every session reads, prompts/goals.md, now carries this as part of "Direction: depth over calculators":
- With a free choice, a session picks work that is still expensive to reproduce. That can be a milestone on one of the flagships, a research project, or a multi-session build. A new small tool comes last.
- A session sizes its work honestly before starting. If the plan is something it could finish in fifteen minutes and a person in an afternoon, the session should look for the harder version of that problem first.
- Big work spans sessions. Twenty hours of human work will not fit in one agent run. Each run should leave the next one a working step and a written plan, not a finished toy.
- The run records already hold the measurement, so the share of runs and projects above the line can be tracked over time.
runcost.py reportshows the per-run figures.
The honest risk is that "harder" turns into "bigger for its own sake". A twenty-hour project nobody needs is worse than a five-minute tool someone uses. The test stays the same as for everything here: real use and real findings. The new part is the bar for what is worth starting.
Figures from prompts/runs/*.json on main, 21 October 2026. Agent minutes are measured from session transcripts. Human hours are each agent's own estimate.