ThoughtsJul 3, 20264 min read
RTFM: Read the F*cking Manual
We finally found an agent that does, and the reading paid its way. On the curiosity comparison.
A companion to the research note Curiosity Comparison Between Agents. Every figure below is that study’s; this piece argues about them rather than adding to them.
Two studies in a row said the same discouraging thing: we build agents a better room and they do not look around it. Then a third agent walked in and read the manual.
Gemini joined the same context ladder the other two had climbed, and broke the pattern. It tops the curiosity index at 100, against Claude's 60 and Codex's 23. It spends 38 percent of its actions reading files. And it is the only one of the three that explores more as the workspace gets richer, rather than reading a fixed amount and stopping.
That last property is the one that matters, and it is easy to skate past. An agent that reads a constant amount cannot be helped by a better environment; you are authoring into a fixed-size window. An agent whose reading scales with what is there is one that can actually collect on the work you put into its workspace.
And it was not paying extra for the privilege
The obvious objection is that thoroughness is just expensive: of course the agent that opens everything burns more. It did not. Gemini averaged about 144K tokens per run against Claude's roughly 164K, while reading far more of the workspace. Codex spent the least of the three, but that is the frugality of an agent that mostly is not looking.
So the extra curiosity was not bought with a bloated bill. Reading turned out to be cheap relative to flailing.
Why this is the hopeful one
The worry running through the earlier studies was structural. If consumption is the bottleneck rather than authoring, then a perfect memory nobody opens is worth nothing, and no amount of better environment design fixes an agent that will not read. That is a bleak conclusion for anyone investing in workspaces.
Gemini is a counterexample worth taking seriously. If the leverage really is in the environment, you want the agent that reaches for it, and of the three tested, that agent exists. The note is careful that this is one probe against three agents, not a ranking that will hold; none of the three mined the deepest seeded memory.
The curiosity index, the per-agent action breakdown and the honest reading are in the research note.