6 comparisons, one per thing worth comparing

Each of these is one page answering one question. Reading files and running shell commands are universal, so what separates the implementations there stays invisible until it costs you something — an indexing base, a missing staleness check. A browser or a memory store is a real choice, and far fewer harnesses ship one.

Nine harnesses read files. What actually differs?6/9 will let an agent overwrite a file that changed after it was readRead13 tools9 read from source30 dimensions13/13 harnesses ship a tool for thisI want my agent to drive a browser. Which one?Browser3 toolsinventorybenchmark declared3/13 harnesses ship a tool for thisWhich harnesses can search the web without an add-on?Web search10 toolsinventory9/13 harnesses ship a tool for thisHow does each harness find code it has not read yet?Code search25 tools20 read from source16 dimensions13/13 harnesses ship a tool for thisEvery harness runs shell commands. What differs underneath?Shell24 toolsinventory13/13 harnesses ship a tool for thisWhich agents remember anything between sessions?Long-term memory5 toolsinventory4/13 harnesses ship a tool for this

One column per harness, 13 of them, in the same order on every rowfaded — reached only through another toolno tool for it

A capability here is the unit of a comparison page, not a slot in a taxonomy: the test for adding one is whether you can write the question its page answers. 160 of the 237 tools on file carry no capability at all — todo lists, subagent spawning, approval flows. They stay ordinary tools until there is a comparison worth reading.