vidtheque.

read-only demo

Videos LqLoYksJ6do

We Gave an Agent Production Code Access and Then Tried to Sleep at Night — Moritz Johner, Form3

index_state ready data_status ok

AI Engineer· published 2026-07-20· 0:21:56· en-US· indexed 2026-08-10 19:46

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

51 shot(s).

keyframes kept every frame deduplicated

What was stored

cues
263
whisperx 263
chunks
39
from 263 cues
keyframes
15
kept of 51 captured
frames with text
15
263 lines read
chapters
0
from the source metadata
keyframe bytes
5.0 MB
word timings on 263 cues

Provenance

Each pipeline stage, its state and the model that produced it
stage state model started took
fetch done — 2026-08-10 03:51 1m 52s
stt done — 2026-08-10 03:53 23s
chunk done — 2026-08-10 03:53 0s
text_embed done — 2026-08-10 19:46 1s
keyframe done — 2026-08-10 03:53 2m 05s
ocr done — 2026-08-10 03:55 5s
frame_embed done — 2026-08-10 19:46 2s

Frames, and what the machine read

  • 0:02 #0 done2 line(s)

    shot 0·sharpness 453.4

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

  • 0:03 #1 done2 line(s)

    shot 1·sharpness 665.7

    1. AlEngineer0.95
    2. World's Fair0.99

  • 0:10 #2 done24 line(s)

    shot 2·sharpness 2743.2

    1. LAB & PLATINUM SPONSORS0.98
    2. Amazon AGI Lab0.98
    3. ANTHROP\C1.00
    4. Google DeepMind1.00
    5. MINIMAX0.94
    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. :neo4j0.93
    15. ORACLE1.00
    16. PayPal1.00
    17. qodo1.00
    18. reducto1.00
    19. Sonar1.00
    20. Makers of0.99
    21. togetherai1.00
    22. Unblocked1.00
    23. WorkOS1.00
    24. SonarQube1.00

  • 0:38 #3 done9 line(s)

    shot 3·sharpness 2463.8

    1. AlEngineer0.97
    2. World'sFair0.96
    3. PRESENTED BY1.00
    4. PITCH0.64
    5. Microsoft1.00
    6. We Gave an Agent Production Code Access1.00
    7. and Then Tried to Sleep at Night0.99
    8. Engineering the future of Al1.00
    9. World'sFair1.00

  • 1:01 #4 done9 line(s)

    shot 4·sharpness 1893.6

    1. AlEngineer0.97
    2. World'sFair0.97
    3. PRESENTED BY1.00
    4. Dependabot exists. Renovate exists.1.00
    5. Microsoft1.00
    6. Problem solved.1.00
    7. TRACK 5· JUNE 30, 20260.95
    8. Security1.00
    9. World'sFair1.00

  • 1:43 #5 done8 line(s)

    shot 5·sharpness 1880.9

    1. AlEngineer0.98
    2. World's Fair0.96
    3. Problem11.00
    4. They patch the manifest.1.00
    5. Everything else is invisible.1.00
    6. TRACK 5· JUNE 30, 20260.96
    7. Security1.00
    8. World'sFair0.98

  • 1:49 #6 done8 line(s)

    shot 6·sharpness 1889.8

    1. AlEngineer0.97
    2. World'sFair0.96
    3. Problem11.00
    4. They patch the manifest.1.00
    5. Everything else is invisible.1.00
    6. TRACK 5· JUNE 30, 20260.96
    7. Security1.00
    8. World'sFair1.00

  • 2:27 #7 skipped

    shot 7·duplicate of #6

  • 3:00 #8 skipped

    shot 8·duplicate of #4

  • 3:25 #9 done7 line(s)

    shot 9·sharpness 2109.3

    1. AlEngineer0.96
    2. World'sFair1.00
    3. A useful coding agent is a supply chain0.98
    4. actor. Whether you planned for that or not1.00
    5. TRACK 5· JUNE 30, 20260.95
    6. World'sFair1.00
    7. Security1.00

  • 3:36 #10 skipped

    shot 10·duplicate of #4

  • 3:57 #11 skipped

    shot 11·duplicate of #4

  • 4:42 #12 skipped

    shot 12·duplicate of #9

  • 5:02 #13 skipped

    shot 13·duplicate of #9

  • 5:30 #14 skipped

    shot 14·duplicate of #4

  • 5:56 #15 done20 line(s)

    shot 15·sharpness 1477.2

    1. AlEngineer0.99
    2. World's Fair0.99
    3. PRESENTED BY1.00
    4. assessment1.00
    5. read repo1.00
    6. cve-remediation1.00
    7. invoke1.00
    8. agent1.00
    9. create PR1.00
    10. + watch Cl0.97
    11. PR remediation1.00
    12. if needed1.00
    13. invoke1.00
    14. agent1.00
    15. post PR comment1.00
    16. send slack notifications1.00
    17. Microsoft1.00
    18. TRACK 5· JUNE 30, 20260.97
    19. Security1.00
    20. World's Fair0.99

  • 6:12 #16 skipped

    shot 16·duplicate of #4

  • 6:31 #17 skipped

    shot 17·duplicate of #9

  • 7:07 #18 skipped

    shot 18·duplicate of #4

  • 7:38 #19 skipped

    shot 19·duplicate of #4

  • 7:56 #20 skipped

    shot 20·duplicate of #9

  • 8:15 #21 skipped

    shot 21·duplicate of #9

  • 8:48 #22 skipped

    shot 22·duplicate of #9

  • 9:24 #23 skipped

    shot 23·duplicate of #4

Transcript

263 cues· 3,731 words· 19,875 chars

  1. 0:13 Thanks, everyone, for joining in.
  2. 0:14 Thanks for the great intro, by the way.
  3. 0:17 So my talk today is about some titles we gave on Azure production code access and then tried to sleep at night.
  4. 0:24 So it's mostly around dependency patching, which is probably the most glamorous problem in software engineering.
  5. 0:33 I guess everyone did it here at some point in the past.
  6. 0:37 It really is like vacuuming like everyone loves it kind of way to do it again next week now at our scale we have thousands of repositories and it really is a backlog that never empties and Because 10 inches today and you know next week 20 more will arrive and you have to deal with them So naturally you think sure Which is automated depend about exists renovate exists problem solved, right?
  7. 1:03 It isn't
  8. 1:04 Because these tools were really built for a world where fixing a CVE means looking at the manifest in a repository and just bumping a version to the next version, and that's it.
  9. 1:17 But our world is much more complicated than that, at least nowadays, since we have containers.
  10. 1:22 So the first problem is that the vulnerable thing isn't necessarily the thing that these tools can see.
  11. 1:28 For instance, the CVE might live in an OS package that you use in your base image.
  12. 1:32 It's not in your Docker file.
  13. 1:34 It's just in the base image.
  14. 1:36 Or if your Docker file pulls down a binary during the build process, it may see the URL, which might contain a version number, but it has no idea how to act on it.
  15. 1:45 So they can only look at the manifest, patch the manifest, but everything else is just simply invisible.
  16. 1:52 Now, the second problem is patches never happen in isolation.
  17. 1:57 So when we look at the Go code base, sometimes you want to bump the Go runtime to the next minor version.
  18. 2:02 When you do that, sometimes you might have to also bump the Go linter, because the versions are just intertwined.
  19. 2:08 You have to bump both of them at the same time.
  20. 2:11 So when you bump the linter, that sometimes introduces new linting rules, which then invalidates the code base.
  21. 2:16 And you're left with this big mess.
  22. 2:19 Now, these old tools, quote unquote old tools, Dependabot and Renovatebot, well, they just bump the Go runtime version, create a PR, and walk away.
  23. 2:27 And you're just left with this mess.
  24. 2:28 And you need to figure out how to fix that.
  25. 2:36 So you don't really have a patching problem.
  26. 2:39 You also have a reasoning problem that you need to address here.
  27. 2:44 You really want to have a tool that looks at the CV and understands the full surface of it.
  28. 2:48 Where does the CV actually live?
  29. 2:50 And it needs to figure out what else needs to move in order to get to a green CI.
  30. 2:57 So that's basically what we did.
  31. 2:59 We built Patch Pilot.
  32. 3:01 We pushed the production.
  33. 3:03 Eventually, InfoSec came around the corner and asked a very reasonable question.
  34. 3:07 Is this automation, or is it a supply chain that's been waiting to happen?
  35. 3:12 A useful coding agent is a supply chain actor, whether you plan for that or not.
  36. 3:17 That's the thesis of this talk, basically.
  37. 3:20 It's not agents are dangerous or agents are fine.
  38. 3:22 It's the moment where you give an agent production credentials in order to be useful.
  39. 3:29 It really becomes a supply chain actor, just like an engineer in your department.
  40. 3:33 And you should apply similar or the same guardrails for the agent, just as you do for engineers.
  41. 3:41 Now, this is a case study.
  42. 3:41 We built this.
  43. 3:42 We ran it in production.
  44. 3:43 InfoSec pushed back in several places.
  45. 3:45 They were right.
  46. 3:46 I just want to present to you what we came up with in the end.
  47. 3:50 It's nothing new.
  48. 3:50 It's nothing fancy.
  49. 3:52 I just want to share what we did and then hopefully have a discussion afterwards to see how we all figure this out together.
  50. 4:01 Now, PatchPilot has two layers.

Open at this second