Which tool should your agent use?

Featured

13 agents, 6 capabilities, who has what

Grey is absent. Codex has no read tool and goes through the shell; half the harnesses have no long-term memory at all. This table is the floor every comparison page is built on.

235 tools on file · every row read from source

claude-code 1→import fs

codeximport fs

hermes1|import fs

mimo-code1: import fs

Comparison · read

Nine read tools, four line-number formats

Two harnesses guard against the model writing the prefix back into your source file. They picked two different places to do it, and neither has data on how often it happens.

Read from source, September 2026

6/9

Finding

Most read tools will let you overwrite someone else's edit

If the file changes after the agent reads it, the write goes through anyway. You find out when the data is already gone.

An invariant, not a preference

1/8

Finding

One read tool in nine counts from zero

Qwen Code's read_file is 0-indexed; everything else that answers the question is 1-indexed. Nothing surfaces this until a line is off by one.

A design choice, recorded not judged