What agent performance actually measures, and why it is only worth 11 points
Autonomy is the most impressive axis to demo and the third most important to the people voting. Here is why the community keeps ranking it below design and speed.
By Owen Castellan
Agent performance is worth 11 of the 100 points on this board. It measures how far a tool gets on a multi step brief without a human intervening, and how well it recovers from its own mistakes.
It is also the axis most likely to produce an impressive demo, which is why it is worth explaining the gap between demo impact and vote weight.
Variance is the tax on autonomy
Our panel ran a forty minute unattended brief on every tool that claims long horizon autonomy. The best result was genuinely striking: real progress on backend logic that other builders skipped entirely.
The second run on the same tool misread the brief in the first three minutes and then worked confidently for thirty seven more. The cost of a wrong plan scales with how long the agent is willing to keep going.
Why builders discount it
Read the comments on agent performance votes and the same calculation appears repeatedly. Autonomy is worth a lot when the plan is right and negative when it is wrong, and builders do not yet trust their ability to tell which they are getting before they pay for it.
Design and speed have no equivalent downside. A fast tool is fast on the bad runs too.
What would change the weight
The weights are set from community voting behaviour, not from our opinion, and they are reviewed monthly. The thing that would move agent performance up is not more autonomy, it is earlier failure. A tool that stops and asks after three minutes when the plan is wrong would be worth more to voters than one that runs for forty and is occasionally brilliant.