Add ClawBench to agent-specific evaluation

[reacher-z]

Summary

Adds one annotated ClawBench entry to section 9, beside live-web and trajectory-evaluation benchmarks.

Why it clears the inclusion bar

• open-source code and an accompanying arXiv paper
• 153 consequential live-site tasks spanning 144 websites and 15 life domains
• isolated task containers and support for multiple agent harnesses
• replay, action, network, and message traces that support both outcome and trajectory analysis

The entry follows the repository’s Name — Org — URL · type — one-line why format and marks this 2026 work as new.

Disclosure: I am a ClawBench maintainer submitting this entry for consideration.

Project links: Website · GitHub · Paper

[xdotli]

Closing this one in favor of #64, which covers the same paper — but to be clear about which parts of each were right, since it's a split decision:

#54 got the section right. §9 · Agent-specific evaluation is where ClawBench belongs; #64 placed it in §6 and I've asked for it to move here.

#64 got the URL right. github.com/reacher-z/ClawBench 301-redirects to TIGER-AI-Lab/ClawBench, so the link in this PR is a stale personal alias.

One correction worth carrying over: this entry says the tasks run in "isolated task containers," but the abstract makes the opposite point — that ClawBench operates on production websites rather than offline sandboxes, with a submission-interception layer as the safeguard. That distinction is the benchmark's main selling point, so it's worth stating accurately in #64.

Also: this is the fifth PR for ClawBench (#56, #59, #60 self-closed, plus #54 and #64 both open simultaneously). All good — but going forward, one open PR per resource makes review much easier, and #64 is the one to iterate on.