vidtheque.

read-only demo

Videos 9QebvrrY3KY

Claude for Long-Horizon Tasks — Lance Martin, Anthropic

index_state ready data_status ok

AI Engineer· published 2026-07-22· 0:25:18· en-US· indexed 2026-08-10 19:43

Open on YouTube

Scene timeline

  1. Shot 0, 0:00 to 0:03, 1 of 1 keyframes kept
  2. Shot 1, 0:03 to 0:05, 1 of 1 keyframes kept
  3. Shot 2, 0:05 to 0:12, 1 of 1 keyframes kept
  4. Shot 3, 0:12 to 0:15, 1 of 1 keyframes kept
  5. Shot 4, 0:15 to 0:34, 1 of 1 keyframes kept
  6. Shot 5, 0:34 to 0:54, 1 of 1 keyframes kept
  7. Shot 6, 0:54 to 1:30, 1 of 1 keyframes kept
  8. Shot 7, 1:30 to 2:05, 0 of 1 keyframes kept
  9. Shot 8, 2:05 to 2:31, 0 of 1 keyframes kept
  10. Shot 9, 2:31 to 2:57, 0 of 1 keyframes kept
  11. Shot 10, 2:57 to 3:23, 0 of 1 keyframes kept
  12. Shot 11, 3:23 to 4:12, 0 of 1 keyframes kept
  13. Shot 12, 4:12 to 4:46, 0 of 1 keyframes kept
  14. Shot 13, 4:46 to 5:20, 0 of 1 keyframes kept
  15. Shot 14, 5:20 to 5:49, 1 of 1 keyframes kept
  16. Shot 15, 5:49 to 6:19, 0 of 1 keyframes kept
  17. Shot 16, 6:19 to 6:24, 0 of 1 keyframes kept
  18. Shot 17, 6:24 to 7:01, 1 of 1 keyframes kept
  19. Shot 18, 7:01 to 7:19, 0 of 1 keyframes kept
  20. Shot 19, 7:19 to 7:48, 0 of 1 keyframes kept
  21. Shot 20, 7:48 to 8:17, 0 of 1 keyframes kept
  22. Shot 21, 8:17 to 8:45, 0 of 1 keyframes kept
  23. Shot 22, 8:45 to 9:13, 0 of 1 keyframes kept
  24. Shot 23, 9:13 to 9:40, 0 of 1 keyframes kept
  25. Shot 24, 9:40 to 10:08, 0 of 1 keyframes kept
  26. Shot 25, 10:08 to 10:14, 0 of 1 keyframes kept
  27. Shot 26, 10:14 to 10:59, 1 of 1 keyframes kept
  28. Shot 27, 10:59 to 11:46, 0 of 1 keyframes kept
  29. Shot 28, 11:46 to 12:01, 1 of 1 keyframes kept
  30. Shot 29, 12:01 to 12:26, 1 of 1 keyframes kept
  31. Shot 30, 12:26 to 12:52, 0 of 1 keyframes kept
  32. Shot 31, 12:52 to 13:17, 0 of 1 keyframes kept
  33. Shot 32, 13:17 to 14:01, 0 of 1 keyframes kept
  34. Shot 33, 14:01 to 14:33, 1 of 1 keyframes kept
  35. Shot 34, 14:33 to 15:04, 0 of 1 keyframes kept
  36. Shot 35, 15:04 to 15:36, 0 of 1 keyframes kept
  37. Shot 36, 15:36 to 16:08, 0 of 1 keyframes kept
  38. Shot 37, 16:08 to 16:20, 0 of 1 keyframes kept
  39. Shot 38, 16:20 to 16:45, 1 of 1 keyframes kept
  40. Shot 39, 16:45 to 17:10, 0 of 1 keyframes kept
  41. Shot 40, 17:10 to 17:35, 0 of 1 keyframes kept
  42. Shot 41, 17:35 to 18:01, 0 of 1 keyframes kept
  43. Shot 42, 18:01 to 18:21, 0 of 1 keyframes kept
  44. Shot 43, 18:21 to 18:46, 0 of 1 keyframes kept
  45. Shot 44, 18:46 to 19:11, 0 of 1 keyframes kept
  46. Shot 45, 19:11 to 19:36, 0 of 1 keyframes kept
  47. Shot 46, 19:36 to 20:01, 0 of 1 keyframes kept
  48. Shot 47, 20:01 to 20:05, 0 of 1 keyframes kept
  49. Shot 48, 20:05 to 20:32, 0 of 1 keyframes kept
  50. Shot 49, 20:32 to 21:00, 0 of 1 keyframes kept
  51. Shot 50, 21:00 to 21:27, 0 of 1 keyframes kept
  52. Shot 51, 21:27 to 21:55, 0 of 1 keyframes kept
  53. Shot 52, 21:55 to 22:22, 0 of 1 keyframes kept
  54. Shot 53, 22:22 to 22:49, 0 of 1 keyframes kept
  55. Shot 54, 22:49 to 23:17, 0 of 1 keyframes kept
  56. Shot 55, 23:17 to 23:44, 0 of 1 keyframes kept
  57. Shot 56, 23:44 to 24:11, 0 of 1 keyframes kept
  58. Shot 57, 24:11 to 24:39, 1 of 1 keyframes kept
  59. Shot 58, 24:39 to 25:01, 0 of 1 keyframes kept
  60. Shot 59, 25:01 to 25:17, 0 of 1 keyframes kept
  61. Shot 60, 25:17 to 25:18, 1 of 1 keyframes kept

61 shot(s).

keyframes kept every frame deduplicated

What was stored

cues
282
whisperx 282
chunks
44
from 282 cues
keyframes
16
kept of 61 captured
frames with text
15
239 lines read
chapters
0
from the source metadata
keyframe bytes
6.3 MB
word timings on 282 cues

Provenance

Each pipeline stage, its state and the model that produced it
stage state model started took
fetch done — 2026-08-09 23:50 0s
stt done — 2026-08-09 20:44 27s
chunk done — 2026-08-09 20:45 0s
text_embed done — 2026-08-10 19:43 0s
keyframe done — 2026-08-09 20:45 3m 13s
ocr done — 2026-08-09 20:48 5s
frame_embed done — 2026-08-10 19:43 3s

Frames, and what the machine read

  • 0:02 #0 done2 line(s)

    shot 0·sharpness 451.6

    1. AlEngineer0.96
    2. World's Fair0.97

  • 0:03 #1 done2 line(s)

    shot 1·sharpness 663.0

    1. AIEngineer0.96
    2. World's Fair0.99

  • 0:10 #2 done24 line(s)

    shot 2·sharpness 2728.3

    1. LAB & PLATINUM SPONSORS0.99
    2. Amazon AGI Lab0.97
    3. ANTHROP\C1.00
    4. Google DeepMind1.00
    5. MINIMAX0.95
    6. OpenAI0.92
    7. Akamai1.00
    8. arize1.00
    9. aws1.00
    10. Braintrust bright data0.99
    11. B1.00
    12. Browserbase1.00
    13. docker1.00
    14. :neo4j0.92
    15. ORACLE1.00
    16. PayPal1.00
    17. qodo1.00
    18. reducto1.00
    19. Sonar1.00
    20. Makers of1.00
    21. togetherai1.00
    22. Unblocked1.00
    23. WorkOS1.00
    24. SonarQube1.00

  • 0:15 #3 done12 line(s)

    shot 3·sharpness 762.2

    1. AlEngineer0.99
    2. ANTHROP\C1.00
    3. World'sFair1.00
    4. Towards asynchronous agents1.00
    5. PRESENTEDBY1.00
    6. Microsoft1.00
    7. AlEnginoor0.93
    8. World'sFair1.00
    9. Lance Martin, Member of Technical Staff, Anthropic0.99
    10. @rlancemartin1.00
    11. World'sFair0.97
    12. Engineering the future of Al1.00

  • 0:23 #4 done9 line(s)

    shot 4·sharpness 1773.5

    1. AlEngineer0.98
    2. ANTHROP\C1.00
    3. World'sFair0.98
    4. Towards asynchronous agents1.00
    5. PRESENTEDBY1.00
    6. Microsoft1.00
    7. Lance Martin, Member of Technical Staff, Anthropic0.99
    8. @rlancemartin1.00
    9. Engineering the future of Al1.00

  • 0:49 #5 done8 line(s)

    shot 5·sharpness 1309.6

    1. AlEngineer0.99
    2. World'sFair1.00
    3. PRESENTED BY0.96
    4. Microsoft1.00
    5. ANTHROP\C1.00
    6. World's Fair0.94
    7. TRACK 1· JUNE 30, 20260.97
    8. Claws & Personal Agents0.99

  • 1:18 #6 done32 line(s)

    shot 6·sharpness 1411.7

    1. AlEngineer0.99
    2. Task duration (time for a human)0.99
    3. World'sFair1.00
    4. Autocomplete/Chat1.00
    5. Synchronous local agents1.00
    6. Asynchronous remote agents1.00
    7. 1 day0.97
    8. 12 hr0.99
    9. Claude Opus 4.61.00
    10. 12 hr0.99
    11. Claude Opus 4.51.00
    12. 4 hr0.88
    13. 4.9 hr0.93
    14. PRESENTEDBY1.00
    15. Claude Opus 41.00
    16. 1.7 hr0.99
    17. Microsoft1.00
    18. 1 hr0.97
    19. Claude Sonnet 3.71.00
    20. Claude Sonnet 3.51.00
    21. 21 min0.99
    22. 10 min1.00
    23. Claude 3 Opus1.00
    24. 4 min0.99
    25. 1 min0.98
    26. 20241.00
    27. 20251.00
    28. 20261.00
    29. ANTHROP\C0.99
    30. World's Fair0.90
    31. TRACK 1· JUNE 30, 20260.97
    32. Claws & Personal Agents0.99

  • 1:58 #7 skipped

    shot 7·duplicate of #6

  • 2:14 #8 skipped

    shot 8·duplicate of #6

  • 2:42 #9 skipped

    shot 9·duplicate of #6

  • 3:20 #10 skipped

    shot 10·duplicate of #6

  • 3:29 #11 skipped

    shot 11·duplicate of #6

  • 4:29 #12 skipped

    shot 12·duplicate of #6

  • 5:09 #13 skipped

    shot 13·duplicate of #6

  • 5:23 #14 done9 line(s)

    shot 14·sharpness 1958.5

    1. AlEngineer0.99
    2. World'sFair1.00
    3. Session1.00
    4. PRESENTEDBY1.00
    5. Microsoft1.00
    6. ANTHROP\C1.00
    7. Anthropic Engineering Blog1.00
    8. TRACK 1· JUNE 30, 20260.97
    9. Claws & Personal Agents0.99

  • 5:56 #15 skipped

    shot 15·duplicate of #14

  • 6:21 #16 skipped

    shot 16·duplicate of #6

  • 6:50 #17 done8 line(s)

    shot 17·sharpness 1586.0

    1. AlEngineer0.99
    2. World's Fair0.96
    3. PRESENTEDBY1.00
    4. Microsoft1.00
    5. ANTHROP\C1.00
    6. World's Fair0.93
    7. TRACK 1· JUNE 30, 20260.97
    8. Claws & Personal Agents0.99

  • 7:15 #18 skipped

    shot 18·duplicate of #6

  • 7:31 #19 skipped

    shot 19·duplicate of #6

  • 8:00 #20 skipped

    shot 20·duplicate of #6

  • 8:31 #21 skipped

    shot 21·duplicate of #6

  • 8:56 #22 skipped

    shot 22·duplicate of #6

  • 9:18 #23 skipped

    shot 23·duplicate of #6

Transcript

282 cues· 4,225 words· 23,361 chars

  1. 0:12 Good to go.
  2. 0:14 All right.
  3. 0:14 Well, take a quick sip and then let's start.
  4. 0:19 It is great to be here.
  5. 0:21 This is like my third year coming to this conference and I always really enjoy it.
  6. 0:25 And thank you for coming to this workshop.
  7. 0:26 I know there's many interesting talks.
  8. 0:29 Let me talk a little bit about our view of asynchronous agents at Anthropic and some things we've been up to lately.
  9. 0:36 So this is kind of a way I think about models and product.
  10. 0:39 So you can think about Cloud as a light source, and you can think about products as windows that allow the light to pass through.
  11. 0:46 And what's kind of interesting is over time, the window that you need to actually kind of see the light of the model kind of shifts.
  12. 0:54 And we've seen this over the past few years.
  13. 0:57 So I'm plotting here different Cloud models and their task horizon.
  14. 1:01 So how much autonomous work can they do over time?
  15. 1:05 And you might recall back in the Opus 3 days, this was kind of like 2024, models could only do maybe 10 to 20 minutes of autonomous work.
  16. 1:13 This is measured by meter.
  17. 1:15 And in that regime, only certain product surfaces made sense.
  18. 1:18 Things like autocomplete, things like chat, where your human is very in the loop.
  19. 1:21 Because the model's really only doing a very short amount of work before you're steering it.
  20. 1:26 Now, the past year, we saw the rise of synchronous coding agents like Cloud Code, and this is kind of a shift because then models could do maybe an hour of work.
  21. 1:35 So it made sense to have them run, but typically locally, where you could still steer them easily.
  22. 1:41 And it's kind of interesting because during this regime, I remember efforts, and I was involved in some efforts to build kind of asynchronous agents,
  23. 1:48 But when models can only do like an hour of work, async as an experience is kind of bad.
  24. 1:54 The model goes off and it like hits an error and it comes back to you over a short period of time.
  25. 1:59 In order to really unlock async, we needed longer task horizons.
  26. 2:03 And so we're starting to see that now.
  27. 2:06 And kind of with this shift in capability and time horizon came a shift in the API surfaces.
  28. 2:12 So if you look at the lower left,
  29. 2:15 Message API came out like two years ago.
  30. 2:17 It's basically prompt response.
  31. 2:19 It's great for building harnesses.
  32. 2:22 But it's a very simple API.
  33. 2:23 Again, you're just passing instead of messages, you get a response out.
  34. 2:27 There's no sense of deployment with that.
  35. 2:29 So you basically take messages API and you can roll your own harness.
  36. 2:32 You can deploy that harness and you have an agent.
  37. 2:34 Now, over the past year, we saw the rise of coding agents in particular.
  38. 2:39 So we released Asian SDK.
  39. 2:40 So that's basically a way to programmatically call cloud code.
  40. 2:43 And that's basically us giving you a harness.
  41. 2:47 But over the past few months since April, as we've seen longer and longer task horizons, we released a new API called Managed Agents, which basically packages both the harness as well as all the managed deployment infrastructure for you.
  42. 3:01 And I'll talk about some of the themes that underpin this new surface, Cloud Managed Agents, and some of the themes that
  43. 3:07 kind of extend beyond just managed agents broadly to think about this kind of new type of asynchronous agents, which can apply, of course, to clause and other types of kind of longer running long horizon agents.
  44. 3:19 So theme one is decoupling the brain from the hands.
  45. 3:24 So when we first set out to build managed agents, we started with the container.
  46. 3:28 We put the harness in the sandbox in the same container.
  47. 3:32 Now, the problem here is what happens if the harness dies or the container dies?
  48. 3:39 What we saw is we actually lose the session.
  49. 3:42 So, basically, this architecture is kind of tricky for long horizon agents because what can happen is your agent is running and if that container dies, you lose everything with it.
  50. 3:54 Also, as models get more capable, putting the credentials in the same container with the agent itself can be problematic.

Open at this second