Skip to content

Compare Thermite

Autoresearch tools run the same loop: an agent changes something, a run measures it, and the better version stays. They differ in who you have to trust at each step, and in how much you have to set up. These questions separate them.

QuestionThermiteKarpathy’s autoresearchpi-autoresearchCORAL
Who writes the benchmark?You, in a kit the agents never seeIt is fixed in the repositoryThe agent, which may change it during the loopYou, and agents can read its source
Who reports the number?The judgeThe agent, from a logThe agentThe grader
Who decides what stays?The judge, by the rules in the contractThe agent, by a rule in its promptThe agent, unless the checks failedNothing is dropped; the score ranks every attempt
What happens to a change that loses?Kept, with its code and the reasonsReset; one line stays in a results fileIts code is reverted; the agent’s notes stay in the logKept, with its commit and score
How is noise handled?Paired blocks, then fresh runs to confirmOne runAn advisory scoreOne grade, unless your grader repeats
What ends the work?Only a limit you setOnly youYour limit, or built-in stops at 200 automatic turns and after more than 20 failures in a rowOnly a limit you set
How many agents?Many, each isolated, with one plannerOneOneMany, sharing their work
What does it take to start?Docker or Podman, about twenty minutes, then a kit for each projectuv, one NVIDIA GPU and a promptOne install command; the agent sets up the restOne install script and a grader

Thermite’s authors write these pages, so read them with that in mind. To keep them honest:

  • Each page names the versions it compares and the date it was checked.
  • Each claim about another tool links to that tool’s code or documentation at that version.
  • Each page says where the other tool is the better choice, and what Thermite costs you.

If a page is wrong or out of date, open an issue.