vidtheque.

read-only demo

Videos JnubYCYunk8

Browser Agents Don't Need Better Models. They Need Better Eyes. - Kushan Raj, ARK

index_state ready data_status ok

AI Engineer· published 2026-06-28· 0:04:25· en-US· indexed 2026-08-11 08:22

Open on YouTube

Scene timeline

  1. Shot 0, 0:00 to 0:02, 1 of 1 keyframes kept
  2. Shot 1, 0:02 to 0:04, 1 of 1 keyframes kept
  3. Shot 2, 0:04 to 0:09, 1 of 1 keyframes kept
  4. Shot 3, 0:09 to 0:10, 1 of 1 keyframes kept
  5. Shot 4, 0:10 to 0:20, 1 of 1 keyframes kept
  6. Shot 5, 0:20 to 0:23, 1 of 1 keyframes kept
  7. Shot 6, 0:23 to 0:43, 1 of 1 keyframes kept
  8. Shot 7, 0:43 to 0:44, 1 of 1 keyframes kept
  9. Shot 8, 0:44 to 0:46, 1 of 1 keyframes kept
  10. Shot 9, 0:46 to 0:49, 1 of 1 keyframes kept
  11. Shot 10, 0:49 to 0:54, 1 of 1 keyframes kept
  12. Shot 11, 0:54 to 0:56, 0 of 1 keyframes kept
  13. Shot 12, 0:56 to 1:02, 0 of 1 keyframes kept
  14. Shot 13, 1:02 to 1:10, 1 of 1 keyframes kept
  15. Shot 14, 1:10 to 1:13, 0 of 1 keyframes kept
  16. Shot 15, 1:13 to 1:17, 1 of 1 keyframes kept
  17. Shot 16, 1:17 to 1:20, 1 of 1 keyframes kept
  18. Shot 17, 1:20 to 1:21, 1 of 1 keyframes kept
  19. Shot 18, 1:21 to 1:23, 1 of 1 keyframes kept
  20. Shot 19, 1:23 to 1:25, 0 of 1 keyframes kept
  21. Shot 20, 1:25 to 1:28, 1 of 1 keyframes kept
  22. Shot 21, 1:28 to 1:29, 0 of 1 keyframes kept
  23. Shot 22, 1:29 to 1:31, 1 of 1 keyframes kept
  24. Shot 23, 1:31 to 1:35, 1 of 1 keyframes kept
  25. Shot 24, 1:35 to 1:37, 1 of 1 keyframes kept
  26. Shot 25, 1:37 to 1:38, 1 of 1 keyframes kept
  27. Shot 26, 1:38 to 1:42, 1 of 1 keyframes kept
  28. Shot 27, 1:42 to 1:45, 1 of 1 keyframes kept
  29. Shot 28, 1:45 to 1:50, 1 of 1 keyframes kept
  30. Shot 29, 1:50 to 1:51, 1 of 1 keyframes kept
  31. Shot 30, 1:51 to 1:56, 1 of 1 keyframes kept
  32. Shot 31, 1:56 to 2:03, 1 of 1 keyframes kept
  33. Shot 32, 2:03 to 2:06, 1 of 1 keyframes kept
  34. Shot 33, 2:06 to 2:07, 1 of 1 keyframes kept
  35. Shot 34, 2:07 to 2:14, 1 of 1 keyframes kept
  36. Shot 35, 2:14 to 2:19, 1 of 1 keyframes kept
  37. Shot 36, 2:19 to 2:20, 1 of 1 keyframes kept
  38. Shot 37, 2:20 to 2:22, 1 of 1 keyframes kept
  39. Shot 38, 2:22 to 2:27, 1 of 1 keyframes kept
  40. Shot 39, 2:27 to 2:29, 1 of 1 keyframes kept
  41. Shot 40, 2:29 to 2:33, 1 of 1 keyframes kept
  42. Shot 41, 2:33 to 2:34, 1 of 1 keyframes kept
  43. Shot 42, 2:34 to 2:38, 1 of 1 keyframes kept
  44. Shot 43, 2:38 to 2:41, 0 of 1 keyframes kept
  45. Shot 44, 2:41 to 2:43, 1 of 1 keyframes kept
  46. Shot 45, 2:43 to 2:44, 1 of 1 keyframes kept
  47. Shot 46, 2:44 to 2:46, 1 of 1 keyframes kept
  48. Shot 47, 2:46 to 2:54, 1 of 1 keyframes kept
  49. Shot 48, 2:54 to 2:56, 1 of 1 keyframes kept
  50. Shot 49, 2:56 to 3:00, 1 of 1 keyframes kept
  51. Shot 50, 3:00 to 3:20, 0 of 1 keyframes kept
  52. Shot 51, 3:20 to 3:23, 1 of 1 keyframes kept
  53. Shot 52, 3:23 to 3:28, 1 of 1 keyframes kept
  54. Shot 53, 3:28 to 3:31, 1 of 1 keyframes kept
  55. Shot 54, 3:31 to 3:32, 0 of 1 keyframes kept
  56. Shot 55, 3:32 to 3:33, 1 of 1 keyframes kept
  57. Shot 56, 3:33 to 3:38, 1 of 1 keyframes kept
  58. Shot 57, 3:38 to 3:40, 1 of 1 keyframes kept
  59. Shot 58, 3:40 to 3:41, 1 of 1 keyframes kept
  60. Shot 59, 3:41 to 3:47, 1 of 1 keyframes kept
  61. Shot 60, 3:47 to 3:55, 1 of 1 keyframes kept
  62. Shot 61, 3:55 to 3:57, 1 of 1 keyframes kept
  63. Shot 62, 3:57 to 4:01, 1 of 1 keyframes kept
  64. Shot 63, 4:01 to 4:22, 1 of 1 keyframes kept
  65. Shot 64, 4:22 to 4:25, 1 of 1 keyframes kept

65 shot(s).

keyframes kept every frame deduplicated

What was stored

cues
77
whisperx 77
chunks
8
from 77 cues
keyframes
57
kept of 65 captured
frames with text
57
4,862 lines read
chapters
0
from the source metadata
keyframe bytes
10.1 MB
word timings on 77 cues

Provenance

Each pipeline stage, its state and the model that produced it
stage state model started took
fetch done — 2026-08-11 08:19 1m 09s
stt done — 2026-08-11 08:20 5s
chunk done — 2026-08-11 08:20 0s
text_embed done — 2026-08-11 08:20 0s
keyframe done — 2026-08-11 08:20 37s
ocr done — 2026-08-11 08:21 1m 13s
frame_embed done — 2026-08-11 08:22 10s

Frames, and what the machine read

  • 0:02 #0 done87 line(s)

    shot 0·sharpness 1094.4

    1. demo-video1.00
    2. 品三四品0.67
    3. Q Search0.94
    4. ©Claude-browser challenge0.98
    5. 00.99
    6. 口0.59
    7. Open with QuickTime Player1.00
    8. Rece1.00
    9. Shar0.93
    10. Claude1.00
    11. Browser Navigation Challeng0.99
    12. localhost:51731.00
    13. ☆0.89
    14. 口0.56
    15. Ca0.77
    16. Finish update1.00
    17. vie1.00
    18. vie1.00
    19. A0.97
    20. Appl0.94
    21. Claude1.00
    22. 40.87
    23. x0.75
    24. ..ext file0.85
    25. text file0.86
    26. Desk1.00
    27. Dowr0.95
    28. Opus 4.60.99
    29. D0.93
    30. vie1.00
    31. vie1.00
    32. .ext file0.86
    33. Docu1.00
    34. vie1.00
    35. nage0.96
    36. .exct file0.80
    37. iClou1.00
    38. ...ext file0.85
    39. o...text file0.88
    40. kush1.00
    41. .ext file0.89
    42. .ext file0.89
    43. Kush1.00
    44. HIGH RISK: Claude can take most actions on the1.00
    45. ..text ffile0.85
    46. Maci0.98
    47. Browser Navigation1.00
    48. intemet now. This setting could put your data at0.99
    49. risk. See safe use tios0.97
    50. Maci0.96
    51. Loon0.93
    52. Challenge1.00
    53. element). Complete each step before moving1.00
    54. to the1.00
    55. AirDr0.94
    56. 30 steps. Each step hides a code behind popups, quizzes,0.99
    57. next.1.00
    58. Netw1.00
    59. and distractions. Dismiss the blockers, solve the challenge,1.00
    60. enter the code, and advance.1.00
    61. enter them in input fields when required.1.00
    62. - Look for codes/values on the page and0.99
    63. Bin1.00
    64. Start Challenge1.00
    65. interacting with content behind them.1.00
    66. -Dismiss overlays/popups/modals before1.00
    67. O0.67
    68. Impo0.99
    69. - If something is blocking the page, find the0.99
    70. close/dismiss/accept button.1.00
    71. - After entering a value and clicking0.99
    72. submit/next, wait briefly for the page to1.00
    73. update before proceeding.0.99
    74. - Plan your actions carefully. Prefer doing0.99
    75. things in sequence rather than rushing.0.98
    76. - Do not narrate or explain. Just act.0.96
    77. O0.63
    78. *This Claude with Chrome0.99
    79. O0.59
    80. FIN A0.99
    81. II0.76
    82. 白0.64
    83. 00:031.00
    84. trying to navigate a browser challenge0.98
    85. O0.63
    86. mkt 20.92
    87. All Tags1.00

  • 0:04 #1 done8 line(s)

    shot 1·sharpness 2160.9

    1. Quick Note:1.00
    2. This video have been sped up1.00
    3. to save time1.00
    4. Timestamps:1.00
    5. 00:12 -> intro to the problem0.99
    6. 00:46 -> small demo + hypothesis0.97
    7. 01:27 -> more real world examples0.98
    8. 03:25 -> implementation details (normal speed)1.00

  • 0:06 #2 done85 line(s)

    shot 2·sharpness 1042.2

    1. demo-video1.00
    2. 品三四品0.63
    3. Q Search0.97
    4. ©Claude-browser challenge0.98
    5. 00.89
    6. 口0.80
    7. ①0.86
    8. Open with QuickTime Player1.00
    9. Rece1.00
    10. Shar0.91
    11. Browser challeng0.97
    12. Browser Navigation Challeng0.98
    13. localhost:51731.00
    14. ☆0.57
    15. 口0.53
    16. Ca0.66
    17. Finish update1.00
    18. vie1.00
    19. vie1.00
    20. A0.97
    21. Appli0.92
    22. 'Claude' started debugging this browser0.98
    23. Cancel1.00
    24. o...text ffile0.87
    25. o...text file0.89
    26. Dowr0.89
    27. ...text file0.90
    28. Desk1.00
    29. Claude1.00
    30. 40.98
    31. x0.78
    32. vie1.00
    33. vie1.00
    34. D0.54
    35. Docu0.97
    36. Opus 4.60.99
    37. vie1.00
    38. nage0.99
    39. iClou0.99
    40. Complete the task below by using1.00
    41. You are a browser automation agent.0.99
    42. 3...eextt file0.85
    43. ..texct file0.87
    44. computer tools to interact with the1.00
    45. ..text file0.90
    46. kush1.00
    47. browser.1.00
    48. .text file0.85
    49. D...ext file0.83
    50. 口0.51
    51. Kush1.00
    52. ## Goal0.98
    53. o...text ffile0.88
    54. Maci0.95
    55. Maci0.99
    56. Browser Navigation1.00
    57. challenge. Quiz hint: the correct1.00
    58. "Complete all steps of the browser1.00
    59. Loom0.88
    60. Challenge0.99
    61. [Letter]-Correct Choice'."0.96
    62. answer always matches 'Option1.00
    63. AirDr0.98
    64. Netw0.96
    65. and distractions. Dismiss the blockers, solve the challenge,1.00
    66. 30 steps. Each step hides a code behind popups, quizzes,1.00
    67. ## Guldelines0.94
    68. Bin1.00
    69. enter the code, and advance.1.00
    70. Working0.95
    71. O0.66
    72. Impo1.00
    73. Start Challenge1.00
    74. Take screenshot1.00
    75. Wallet ompty Add credits0.95
    76. Reply to Caude0.93
    77. O0.61
    78. *This Claude with Chrome0.97
    79. O0.75
    80. FIN A0.95
    81. Stop Claude1.00
    82. trying to navigate a browser challenge0.98
    83. O0.56
    84. mkt 20.97
    85. All Tags0.95

  • 0:09 #3 done1 line(s)

    shot 3·sharpness 380.8

    1. Skipping my rant for the sake of time1.00

  • 0:16 #4 done122 line(s)

    shot 4·sharpness 1366.0

    1. demo-video1.00
    2. 品三四品0.82
    3. ...0.57
    4. Q Search0.93
    5. Rece1.00
    6. Claude-browser challenge1.00
    7. 0□0.50
    8. Open with QuickTime Player0.99
    9. Shar0.88
    10. Browser challeng0.99
    11. Browser Navigation Challeng:0.99
    12. +1.00
    13. vie1.00
    14. localhost:5173/step1?version=21.00
    15. ☆0.93
    16. 口0.56
    17. a0.85
    18. Finish update1.00
    19. vie1.00
    20. vie1.00
    21. A Appli0.88
    22. 'Claude' started debugging this browser0.99
    23. Cancel1.00
    24. ×0.88
    25. o...extt file0.84
    26. o...text ffile0.85
    27. Dowr0.84
    28. o...text file0.90
    29. Desk1.00
    30. Step 1 of 30 - Browser Navigation Challenge1.00
    31. Claude1.00
    32. vie1.00
    33. vie1.00
    34. Docu0.99
    35. Cookie Consent1.00
    36. Opus 4.61.00
    37. vie1.00
    38. nage0.88
    39. Kush1.00
    40. Maci1.00
    41. 00.69
    42. iClou1.00
    43. Maci0.99
    44. kush1.00
    45. Challeng1.00
    46. You are on s1.00
    47. Wait for the0.99
    48. UUB1.00
    49. Alert!1.00
    50. This is a popup message. T0.98
    51. is fake! Look for another wa1.00
    52. Dismiss1.00
    53. We use cookies to enhance your experience. Do you1.00
    54. accept?1.00
    55. Accept1.00
    56. lose.1.00
    57. Decline1.00
    58. Close1.00
    59. ge. Click X to close.0.99
    60. azing deals!1.00
    61. Complete the task below by using1.00
    62. You are a browser automation agent.0.99
    63. computer tools to interact with the1.00
    64. browser.1.00
    65. challenge. Quiz hint: the correct1.00
    66. ## Goal0.95
    67. "Complete all steps of the browser0.99
    68. O...eextt file0.83
    69. ...text file0.89
    70. o...text file0.94
    71. o...ext file0.86
    72. o...ext file0.78
    73. o..texct file0.81
    74. Loon0.98
    75. Dismiss1.00
    76. im, quis nostrud exercitation ullamco0.98
    77. answer always matches 'Option0.99
    78. [Letter]- Correct Choice'."0.96
    79. A0.82
    80. AirDr0.99
    81. Netw1.00
    82. Bin1.00
    83. Code revealed: UUB0.98
    84. Duis aute irure dolor ir0.97
    85. pariatur. Exc-0.95
    86. Section 2: Importal0.99
    87. Alert!1.00
    88. This is a popup message. The close button0.99
    89. nulla1.00
    90. ## Guldelines0.95
    91. War1.00
    92. is fake! Look for another way to close.1.00
    93. Revealing and capturing promotional code0.99
    94. Impo0.99
    95. Scroll Down to Fir0.97
    96. Section 31.00
    97. Sed ut pers1.00
    98. This1.00
    99. Dismiss1.00
    100. mque1.00
    101. 40.86
    102. Take screenshot1.00
    103. laudantium,0.99
    104. is fak1.00
    105. Section 41.00
    106. Dismiss1.00
    107. Nemo enim0.99
    108. lunit sed oula0.86
    109. Wallet empty -Add credits0.94
    110. Please select an option above1.00
    111. Submit & Continue0.97
    112. Reply to Caude0.95
    113. O0.52
    114. *This Claude with Chrome0.98
    115. O FIN A0.96
    116. Section 41.00
    117. This is filler content Keep scrolling to find the navigation bu0.98
    118. Stop Claude0.97
    119. trying to navigate a browser challenge1.00
    120. O0.56
    121. mkt 20.94
    122. All Tags..0.89

  • 0:21 #5 done4 line(s)

    shot 5·sharpness 1951.6

    1. Quick Note:0.99
    2. This browser Challenge was released By Brett Adcock on X1.00
    3. It has been taken down since, so this my clone of the challenge0.99
    4. I will open source this since its a pretty cool challenge1.00

  • 0:39 #6 done121 line(s)

    shot 6·sharpness 1323.2

    1. demo-video1.00
    2. 品三四品0.75
    3. ……0.57
    4. Q Search0.91
    5. Rece1.00
    6. Claude-browser challenge1.00
    7. Open with QuickTime Player0.99
    8. Shar0.92
    9. Browser challeng1.00
    10. Browser Navigation Challeng:0.99
    11. +1.00
    12. vie1.00
    13. localhost:5173/step1?version=21.00
    14. ☆0.94
    15. Ca0.70
    16. Finish update1.00
    17. vie1.00
    18. vie1.00
    19. A Appli0.84
    20. 'Claude' started debugging this browser0.99
    21. Cancel1.00
    22. ×0.86
    23. o...text ffile0.88
    24. o...text file0.94
    25. Dowr0.87
    26. D...text file0.88
    27. Desk1.00
    28. Step 1 of 30 - Browser Navigation Challenge0.99
    29. Claude1.00
    30. 40.96
    31. x0.88
    32. vie1.00
    33. vie1.00
    34. Docu1.00
    35. Opus 4.61.00
    36. vie1.00
    37. nage0.96
    38. Kush1.00
    39. 00.67
    40. iClou0.99
    41. Maci0.99
    42. kush1.00
    43. You are on s0.98
    44. Challeng1.00
    45. Wait for the1.00
    46. UUB1.00
    47. Alert!1.00
    48. is fake! Look for another way to close.0.99
    49. This is a popup message. The close button0.99
    50. Dismiss1.00
    51. on0.82
    52. ise button1.00
    53. lose.1.00
    54. Click here for amazing deals!1.00
    55. This is a popup message. Click X to close.0.99
    56. Close1.00
    57. Complete the task below by using1.00
    58. You are a browser automation agent.0.97
    59. computer tools to interact with the0.98
    60. browser.1.00
    61. ## Goal0.96
    62. "Complete all steps of the browser0.98
    63. O...extt file0.84
    64. D...text file0.85
    65. D...ext file0.89
    66. o...text file0.94
    67. o...text file0.87
    68. ..exct file0.86
    69. Maci0.99
    70. challenge. Quiz hint: the correct1.00
    71. Loo0.97
    72. Code revealed!0.98
    73. Dismiss1.00
    74. im, quis nostrud exercitation ullamco0.98
    75. answer always matches 'Option1.00
    76. [Letter] - Correct Choice'."0.94
    77. A0.69
    78. AirDr1.00
    79. Netw1.00
    80. Code revealed: UUB0.98
    81. Section 2: Importal1.00
    82. Alert!1.00
    83. ## Guldelines0.94
    84. Bin1.00
    85. pariatur. Exc0.97
    86. Duis aute irure dolor ir0.99
    87. This is a popup message. The close button0.99
    88. nulla1.00
    89. o0.58
    90. Impo0.99
    91. Scroll Down to Fir0.97
    92. Section 31.00
    93. Sed ut pers1.00
    94. This1.00
    95. War1.00
    96. is fake! Look for another way to close.1.00
    97. Dismiss1.00
    98. mque1.00
    99. 40.86
    100. Removing popup overlays from webpage1.00
    101. Take screenshot1.00
    102. laudantium,0.98
    103. is fak1.00
    104. Javascript tool1.00
    105. Section 41.00
    106. Dismiss1.00
    107. Nemo enim0.92
    108. lunit sed ouia0.81
    109. Wallet empty·Add credits0.96
    110. Please select an option above0.99
    111. Submit & Continue1.00
    112. Reply to Caude0.94
    113. O0.59
    114. *This Claude with Chrome0.99
    115. O FIN A0.90
    116. Section 40.99
    117. This is filler content. Keep scrolling to find the navigation bu0.99
    118. Stop Claude0.98
    119. trying to navigate a browser challenge0.99
    120. mkt 20.95
    121. ① All Tags..0.83

  • 0:44 #7 done100 line(s)

    shot 7·sharpness 730.5

    1. <0.93
    2. demo-video1.00
    3. 品 三 四0.62
    4. 00.58
    5. ……0.68
    6. Q Search0.94
    7. ①0.84
    8. Recents1.00
    9. Name1.00
    10. Date Modified1.00
    11. Size1.00
    12. Kind1.00
    13. Shared1.00
    14. claude-aadhar1.00
    15. Today at 3:40 PM1.00
    16. 131 MB1.00
    17. QT movie1.00
    18. Claude-browser challenge0.99
    19. Today at 3:14 PM1.00
    20. 368.7 MB0.95
    21. QT movie0.99
    22. Favourites1.00
    23. AApplications0.98
    24. feedback-blocker-recovery.md0.97
    25. feedback-action-log.md1.00
    26. claude-trekking1.00
    27. Today at 3:36 PM0.95
    28. Today at 3:37PM1.00
    29. Today at 3:42PM0.99
    30. 145.7 MB0.99
    31. 4 KB0.99
    32. 3 KB1.00
    33. QT movie1.00
    34. Markdo...text file0.97
    35. Markdo...text file0.94
    36. Downloads1.00
    37. feedback-mutations.md1.00
    38. Today at 3:36 PM0.98
    39. 3 KB0.99
    40. Markdo...text fille0.95
    41. Desktop1.00
    42. our-agent1.00
    43. ours-aadhar1.00
    44. Today at 3:29 PM1.00
    45. Today at 3:55 PM0.99
    46. 227.3 MB1.00
    47. 39.9 MB1.00
    48. QT movie1.00
    49. QT movie1.00
    50. Documents1.00
    51. ours-trekking1.00
    52. Today at 4:03 PM0.98
    53. 197.8 MB1.00
    54. QT movie1.00
    55. Screenshot 2026-04-03 at 12.32.23 AM0.99
    56. Today at 12:32 AM1.00
    57. 618 KB1.00
    58. PNG image0.98
    59. Locations1.00
    60. v5-metacritic-real-world.md1.00
    61. Today at 3:35 PM0.99
    62. 3 KB0.99
    63. Markdo...text file0.98
    64. iCloud Drive0.99
    65. v5-step1-full-unedited.md1.00
    66. v5-step1-overlays-blocking.md1.00
    67. Today at 4:06 PM0.98
    68. Today at 3:35 PM0.99
    69. 5 KB0.99
    70. 2 KB0.98
    71. Markdo...text file0.92
    72. Markdo...text file0.97
    73. kushanraj1.00
    74. v5-step9-deep-blocking.md1.00
    75. Today at 3:37 PM0.97
    76. 3 KB0.95
    77. Markdo...text file0.97
    78. v5-step9-fll-unedited.md0.99
    79. Today at 4:09 PM1.00
    80. 6 KB1.00
    81. Markdo...textxfile0.84
    82. Kushan's Mac0.96
    83. v5-step10-dnd.md1.00
    84. Today at 3:35PM1.00
    85. 2 KB0.97
    86. Markdo...text file0.98
    87. Macintosh HD1.00
    88. Macintosh HD1.00
    89. Loom 0.341.0-...0.96
    90. A0.75
    91. AirDrop1.00
    92. #0.54
    93. Network1.00
    94. Bin1.00
    95. Tags1.00
    96. O Important0.98
    97. 40.96
    98. O FIN ACCT0.97
    99. O mkt 20.94
    100. 0 All Tags...0.91

  • 0:46 #8 done72 line(s)

    shot 8·sharpness 2825.4

    1. <0.91
    2. demo-video1.00
    3. 品 三 四 品0.56
    4. ...0.59
    5. Q Search0.96
    6. ①0.81
    7. Recents1.00
    8. our-agent1.00
    9. 00.86
    10. 口0.86
    11. Open with QuickTime Player1.00
    12. Shared1.00
    13. Google Chrome for Testing0.98
    14. File1.00
    15. Edit0.88
    16. View1.00
    17. History1.00
    18. Bookmarks1.00
    19. Profiles0.98
    20. Tab1.00
    21. Window1.00
    22. Help1.00
    23. o0.74
    24. 00.54
    25. Fri 3 Apr 3:25PM1.00
    26. movie1.00
    27. Browser Navigation Challeng0.99
    28. movie1.00
    29. A Applicati0.95
    30. ~/browser-challenge/.claude/worktrees/expand-tool (orphan-controls-a1.00
    31. -compact-th...) $ uv run -m crawler.main --url "http://localhost:5170.99
    32. --goal "Complete all steps of the browser challenge. Quiz hint: th0.99
    33. K0.58
    34. localhost:5173/step17version=10.99
    35. rkdo..teext file0.84
    36. movie1.00
    37. rkdo...text file0.95
    38. Downloa0.98
    39. correct answer always matches 'Option [Letter] - Correct Choice'."0.99
    40. Step 1 of 30 - Browser Navigation C0.98
    41. rkdo...texxt file0.92
    42. gent-mode raw-0.97
    43. --raw-model cerebras/gpt-oss-120b --headed0.99
    44. movie1.00
    45. ☐0.62
    46. Desktop1.00
    47. Run directory: /Users/kushanraj/browser-challenge/.claude/worktrees/1.00
    48. movie1.00
    49. iCloud D0.98
    50. Docume1.00
    51. pand-tool/runs/master_20260403_0955211.00
    52. Navigating to: http://localhost:5173/1.00
    53. Raw model: cerebras/gpt-oss-120b0.99
    54. Agent mode: raw0.99
    55. Challenge Step 10.98
    56. You are on step 1. Complete the challenges and enter the code to proceed.0.99
    57. G image0.98
    58. movie1.00
    59. rkdo...text fille0.95
    60. rkdo...tet file0.93
    61. rkdo...text file0.96
    62. kushanra0.98
    63. [Page 1] URL: http://localhost:5173/1.00
    64. rkdo..text file0.91
    65. Inspect the DOM to find it, or click here 3 more times to reveal.0.99
    66. rkdo...text file0.92
    67. Kushan's0.99
    68. [TIMING] dom_stability=236ms1.00
    69. html_capture=4ms1.00
    70. extraction=26ms1.00
    71. rkdo...textt file0.94
    72. Macintos0.94

  • 0:48 #9 done

    shot 9·sharpness 3677.3

  • 0:50 #10 done

    shot 10·sharpness 3158.1

  • 0:56 #11 skipped

    shot 11·duplicate of #9

  • 0:57 #12 skipped

    shot 12·duplicate of #10

  • 1:05 #13 done

    shot 13·sharpness 2155.5

  • 1:11 #14 skipped

    shot 14·duplicate of #10

  • 1:16 #15 done

    shot 15·sharpness 3373.9

  • 1:18 #16 done

    shot 16·sharpness 3081.9

  • 1:21 #17 done

    shot 17·sharpness 3090.4

  • 1:22 #18 done

    shot 18·sharpness 3364.7

  • 1:24 #19 skipped

    shot 19·duplicate of #10

  • 1:28 #20 done

    shot 20·sharpness 231.6

  • 1:29 #21 skipped

    shot 21·duplicate of #7

  • 1:31 #22 done

    shot 22·sharpness 2492.9

  • 1:35 #23 done

    shot 23·sharpness 2354.9

The page's on-screen-text budget of 600 lines is spent, so the last cards in this grid list fewer lines than they hold. Narrow the page with ?frames= to read them.

Transcript

77 cues· 919 words· 4,807 chars

  1. 0:00 Okay, hey everyone, I am Kushan.
  2. 0:02 I worked at Cellform as a founding engineer for two years.
  3. 0:04 Let's talk about what I'm interested in right now, and that is browser agents.
  4. 0:07 Browser agents, as an idea, are so cool, right?
  5. 0:09 But browser agents should go crazy, right?
  6. 0:12 I personally have not seen that adoption, and me, myself, I don't use browser agents that much.
  7. 0:15 I've been exploring that for some time.
  8. 0:17 I've been trying to understand why that is.
  9. 0:18 So, on my screen right now, we have the browser challenge.
  10. 0:21 But this is a very interesting benchmark for browser agents because there are so many things that you have to do, long rise in sequencing of your tasks.
  11. 0:29 And this actually reveals why browser agents suck.
  12. 0:31 If you saw at the beginning of the video, the browser, this agent took like maybe 10, 20 seconds just to click the start button.
  13. 0:37 And now we're on step one.
  14. 0:38 There are 30 steps and it has taken so long just to click one button.
  15. 0:42 So enough of this.
  16. 0:43 I want to show you what I've been building.
  17. 0:45 So same website.
  18. 0:47 I've tried to sort of replicate the feeling of seeing what's happening you know you can see what the agent is thinking but as you can see it is so much faster and so much quicker and I'm using a much cheaper model right the hypothesis here is models are pretty smart but it's the infra around them that sucks if you noticed in the video earlier maybe I'll put a screenshot the agent is
  19. 1:04 is trying to debug what's going on.
  20. 1:06 It's trying to click something but it doesn't understand what's going on.
  21. 1:08 So my core thesis here has been give a nice environment for the agent to use, right?
  22. 1:12 So where it can plan long sequences, it can figure out where it failed, what is going on, and it can plan the click correctly.
  23. 1:18 I figured out is a cool representation which compresses the website and lets the agent see the entire page in very few tokens.
  24. 1:25 Now I want to show some actual examples.
  25. 1:28 Let's say I want to download my Aadhaar.
  26. 1:30 So this is Claude trying to do it.
  27. 1:31 So I'd assume that this is very simple for a browser agent.
  28. 1:33 I take a screenshot, you see the button right there, click it.
  29. 1:36 But then what's interesting is that it got stuck after this point.
  30. 1:39 So from 46 seconds until the end of this video, it took a screenshot, it scrolled for some reason, it took a screenshot.
  31. 1:45 Basically, this entire process took two minutes, whereas in my case, in our video,
  32. 1:55 So it just boots, and boom, done.
  33. 1:58 And that's the beauty of a browser agent.
  34. 2:00 Just how quick was that?
  35. 2:02 And I'm using such a cheap model for this.
  36. 2:03 Another interesting example is, so my friends and I are going trekking on Sunday.
  37. 2:09 I was wondering, you know what?
  38. 2:11 Because this site is in Canada, and I am not very fluent in Canada.
  39. 2:14 It took me some time to figure out this website.
  40. 2:16 So I asked Claude, hey, can you book this for me?
  41. 2:19 And by the end of it, it's unable to pick a date.
  42. 2:21 And it's just stuck.
  43. 2:22 This is the video of my agent.
  44. 2:34 You can see it selects it and
  45. 2:39 puts in the date, and boom, done, right?
  46. 2:41 It's so simple and convenient to use in theory.
  47. 2:44 So what's next, right?
  48. 2:45 What am I planning on doing?
  49. 2:46 I'm thinking of open sourcing this project because, again, this code is not super defensible.
  50. 2:50 The product that I want to give is, again, maybe an API that, as you can see, we were running these commands.

Open at this second