WEBVTT

1
00:00:00.360 --> 00:00:03.460
Imagine asking an agent to apply for a job using your resume.

2
00:00:04.019 --> 00:00:08.960
It finds the vacancy, extracts your details, uploads a file, and
completes several forms.

3
00:00:09.560 --> 00:00:14.600
But a finished-looking page cannot tell you whether the actual
submission contains the right information.

4
00:00:15.200 --> 00:00:19.120
ClawBench studies this kind of everyday task on live, production
websites.

5
00:00:19.600 --> 00:00:23.900
The central question is simple: can an agent carry your goal
through the entire workflow?

6
00:00:24.480 --> 00:00:28.160
We will follow an illustrative application, then see how the
benchmark checks the evidence.

7
00:00:28.733 --> 00:00:34.733
The local paper contains one hundred fifty-three tasks across
one hundred forty-four platforms and fifteen categories.

8
00:00:35.353 --> 00:00:41.513
These include purchases, reservations, applications, and other
workflows that would change information on a server.

9
00:00:42.333 --> 00:00:47.633
Each task has a user instruction, a starting address, and a
specified final submission request.

10
00:00:48.453 --> 00:00:53.633
Human annotators complete the task and record a reference
trajectory, including its submission payload.

11
00:00:54.433 --> 00:00:59.153
They filter out unavailable sites, paid subscriptions, and
geographically restricted services,

12
00:00:59.513 --> 00:01:00.713
and check reproducibility.

13
00:01:01.513 --> 00:01:07.733
For our example, the goal becomes specific: the right
application, the right resume, and correctly filled fields.

14
00:01:08.333 --> 00:01:11.253
A real submission could place an order or send an application.

15
00:01:11.773 --> 00:01:15.013
ClawBench therefore intercepts the designated final request.

16
00:01:15.493 --> 00:01:20.773
A human identifies its endpoint, request method, and required
payload fields during the reference run.

17
00:01:21.493 --> 00:01:27.573
The instrumentation captures the matching request body, blocks
it before transmission, and logs it for evaluation.

18
00:01:28.253 --> 00:01:33.333
Other page loads and dynamic requests pass through, preserving
ordinary interaction with the live website.

19
00:01:34.113 --> 00:01:39.033
In our example, the agent can reach Submit; the application
request is captured instead of sent.

20
00:01:39.693 --> 00:01:42.233
This boundary keeps the evidence of intended submission

21
00:01:42.553 --> 00:01:45.693
while preventing that annotated transaction from reaching the
server.

22
00:01:46.267 --> 00:01:51.387
The framework records five synchronized evidence streams, rather
than relying on one final screenshot.

23
00:01:52.087 --> 00:01:55.907
Session video and action screenshots show the visible page and
how it changes.

24
00:01:56.587 --> 00:02:03.067
HTTP traffic shows outgoing request data; browser actions record
clicks, typing, scrolling, and navigation.

25
00:02:03.847 --> 00:02:08.507
Agent messages preserve the framework's reasoning traces, tool
calls, and intermediate outputs.

26
00:02:09.327 --> 00:02:14.247
Human reference runs use the same recording infrastructure,
making comparison across the streams possible.

27
00:02:14.947 --> 00:02:19.427
Together, these records let a reviewer connect what was seen,
attempted, and submitted.

28
00:02:19.933 --> 00:02:25.633
A Claude Code sub-agent evaluates the task under a fixed rubric,
using both recorded trajectories.

29
00:02:26.273 --> 00:02:32.193
It aligns corresponding steps, checks required fields and
actions, and compares the terminal submission state.

30
00:02:33.013 --> 00:02:36.333
Consider our application: the uploaded resume is correct,

31
00:02:36.733 --> 00:02:39.673
but the outgoing phone-number field contains a different number.

32
00:02:40.313 --> 00:02:45.773
The payload exposes that mismatch even if the screenshot looked
complete or the agent claimed success.

33
00:02:46.733 --> 00:02:53.893
Interception alone is therefore insufficient. Required actions
and information must also satisfy the evaluation rules.

34
00:02:54.733 --> 00:02:59.793
The evaluator returns a binary verdict with structured
justification grounded in the recorded evidence.

35
00:03:00.400 --> 00:03:06.280
The rubric must also handle borderline cases, including blocked
submissions and equivalent alternative paths.

36
00:03:06.980 --> 00:03:10.660
The authors refine the prompt against calibration examples
carrying human verdicts.

37
00:03:11.200 --> 00:03:14.620
They evaluate the whole set, then collect disagreements with
those human labels.

38
00:03:15.280 --> 00:03:19.160
Each disagreement triggers diagnosis and targeted rule edits to
correct that case.

39
00:03:19.740 --> 00:03:24.600
Next, they retest every calibration example, exposing
regressions introduced by the local changes.

40
00:03:25.120 --> 00:03:30.340
The resulting prompt governs later evaluation: specific repairs,
checked across the whole calibration set.

41
00:03:30.900 --> 00:03:32.160
Now the measured outcome.

42
00:03:32.760 --> 00:03:37.680
The local experiment evaluates six models through the OpenClaw
framework, controlling a Chromium browser.

43
00:03:38.480 --> 00:03:42.840
Claude Sonnet four point six leads at thirty-three point three
percent task success.

44
00:03:43.500 --> 00:03:48.580
GLM-five reaches twenty-four point two percent; Gemini three
Flash reaches nineteen percent.

45
00:03:49.140 --> 00:03:55.300
Claude Haiku four point five scores eighteen point three
percent, while GPT-five point four reaches six point five.

46
00:03:56.020 --> 00:03:59.060
Gemini three point one Flash Lite scores three point three
percent.

47
00:03:59.820 --> 00:04:05.240
These are success rates for this paper's task set and execution
setup, with completion judged against human references.

48
00:04:06.007 --> 00:04:10.227
The recordings also explain where the runs break down, rather
than stopping at a leaderboard.

49
00:04:10.947 --> 00:04:17.447
Across nine hundred eighteen task-model runs, six hundred
nineteen are classified as blocked by anti-bot systems.

50
00:04:17.887 --> 00:04:22.387
That is sixty-seven point four percent of all runs, including
successful ones in the denominator.

51
00:04:23.167 --> 00:04:29.227
This makes live-site access a major measured obstacle in this
setup; the category does not isolate reasoning ability.

52
00:04:30.067 --> 00:04:35.547
Other recorded outcomes include policy refusals, website errors,
incomplete work, and repetitive loops.

53
00:04:36.287 --> 00:04:41.447
For developers, the useful question becomes: which step failed,
and what does its evidence actually show?

54
00:04:42.053 --> 00:04:49.073
ClawBench combines everyday tasks, targeted interception, human
demonstrations, and traceable evaluation.

55
00:04:49.953 --> 00:04:54.653
Our example separates a plausible application screen from a
correctly assembled submission.

56
00:04:55.333 --> 00:04:58.773
The final action is examined without sending the annotated
transaction.

57
00:04:59.893 --> 00:05:04.533
Reported success measures the model, browser framework, and live
environment together.

58
00:05:05.233 --> 00:05:09.093
The research direction is to make useful assistance observable,
step by step.

59
00:05:09.933 --> 00:05:14.113
Confidence comes from completed work, supported by evidence you
can inspect.