vidtheque.

read-only demo

Videos BEKc4P87XKo

Agentic Engineering: Working With AI, Not Just Using It — Brendan O'Leary

index_state ready data_status ok

AI Engineer· published 2026-04-07· 0:27:03· en-US· indexed 2026-08-10 19:46

Open on YouTube

Scene timeline

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

56 shot(s).

keyframes kept every frame deduplicated

What was stored

cues
285
whisperx 285
chunks
49
from 285 cues
keyframes
27
kept of 56 captured
frames with text
27
355 lines read
chapters
0
from the source metadata
keyframe bytes
5.2 MB
word timings on 285 cues

Provenance

Each pipeline stage, its state and the model that produced it
stage state model started took
fetch done — 2026-08-10 01:40 1m 30s
stt done — 2026-08-10 01:41 30s
chunk done — 2026-08-10 01:42 0s
text_embed done — 2026-08-10 19:46 1s
keyframe done — 2026-08-10 01:42 2m 36s
ocr done — 2026-08-10 01:44 14s
frame_embed done — 2026-08-10 19:46 5s

Frames, and what the machine read

  • 0:18 #0 done2 line(s)

    shot 0·sharpness 301.6

    1. Agentic Engineering1.00
    2. Working With AI, Not Just Using It0.99

  • 0:57 #1 skipped

    shot 1·duplicate of #0

  • 1:05 #2 done2 line(s)

    shot 2·sharpness 319.7

    1. The Paradigm Shift1.00
    2. From autocomplete to teammates1.00

  • 1:31 #3 done13 line(s)

    shot 3·sharpness 621.9

    1. The Paradigm Shift1.00
    2. From autocomplete to teammates1.00
    3. 20201.00
    4. 20221.00
    5. 2025+1.00
    6. Autocomplete1.00
    7. Copilots1.00
    8. Agents1.00
    9. Finish your line0.99
    10. Suggest functions1.00
    11. Execute tasks1.00
    12. 90% of tech workers now use AI at work0.99
    13. - Google DORA Report0.99

  • 2:14 #4 skipped

    shot 4·duplicate of #3

  • 2:42 #5 done5 line(s)

    shot 5·sharpness 575.3

    1. The mental model shift:1.00
    2. AI as an energetic, enthusiastic junior developer1.00
    3. "We are no longer just using machines, we are now working with1.00
    4. them.0.99
    5. - Armin Ronacher, creator of Flask0.99

  • 3:15 #6 done14 line(s)

    shot 6·sharpness 806.7

    1. The Agent-Model Architecture1.00
    2. Three-way collaboration: You + Agent + Model1.00
    3. You1.00
    4. Agent1.00
    5. Model1.00
    6. Direction & judgment0.99
    7. Runs locally in your IDE0.98
    8. Runs in the cloud1.00
    9. The thinking1.00
    10. Kilo Code, Cursor, Claude Code0.98
    11. Claude, GPT, Gemini0.97
    12. A0.99
    13. Agents are an illusion - models are stateless0.99
    14. The only way to get better output is to put better tokens in1.00

  • 3:47 #7 skipped

    shot 7·duplicate of #6

  • 4:23 #8 skipped

    shot 8·duplicate of #6

  • 4:42 #9 done5 line(s)

    shot 9·sharpness 732.6

    1. Context Engineering0.98
    2. What you put in determines what you get out - and1.00
    3. it costs more than you think1.00
    4. "Context engineering is the delicate art and science of filling the context window0.99
    5. with just the right information for the next step." - Andrej Karpathy0.99

  • 5:15 #10 done16 line(s)

    shot 10·sharpness 735.9

    1. Problem 1: Context is expensive1.00
    2. Every token you add to the context window costs money - and the cost scales fast.0.99
    3. 10K1.00
    4. 200K1.00
    5. 600K1.00
    6. 1M+1.00
    7. tokens1.00
    8. tokens1.00
    9. tokens1.00
    10. tokens1.00
    11. ~$0.030.99
    12. ~$0.601.00
    13. $1.50+1.00
    14. More context ≠ better results0.98
    15. You're paying for noise, not signal1.00
    16. And even if cost weren't an issue - more context actually makes the model worse.0.99

  • 5:43 #11 done18 line(s)

    shot 11·sharpness 1102.8

    1. Problem 2: More context makes the model0.99
    2. dumber1.00
    3. It's not just about money. Quality degrades well before you hit the context limit.0.99
    4. Smart Zone I Dumb Zone →0.98
    5. 0%1.00
    6. ~50%1.00
    7. 100%1.00
    8. The MCP Trap1.00
    9. The Benchmark1.00
    10. Loading every MCP server you have means0.99
    11. Stay under 50% context utilization when0.99
    12. you're starting every task already in the0.99
    13. possible. The model gets worse and your costs1.00
    14. dumb zone - before you've typed a single0.99
    15. get worse as you approach the limit.1.00
    16. word.1.00
    17. - Dex Horthy0.98
    18. But context bloat isn't the only risk. Sometimes the problem isn't quantity - it's qu0.99

  • 6:07 #12 done19 line(s)

    shot 12·sharpness 1037.7

    1. Problem 3: Bad context poisons everything1.00
    2. Once bad information enters the context window, the model treats it as ground truth - and0.99
    3. builds on it.0.98
    4. 萬0.73
    5. Common sources1.00
    6. The fix0.97
    7. Task mixing1.00
    8. Leftover context from a previous, unrelated task0.99
    9. Outdated comments0.99
    10. Stale docs and ToD0s that no longer reflect reality0.99
    11. Start a new session1.00
    12. You can't think your way out of a poisoned1.00
    13. Massive debug logs1.00
    14. context. Don't try to fix it - reset it.0.99
    15. Thousands of lines of noise drowning out the signal1.00
    16. Compounding hallucinations1.00
    17. An early AI mistake that every subsequent response0.99
    18. builds on1.00
    19. So context is expensive, it degrades quality, and bad context corrupts outputs. How do you1.00

  • 6:40 #13 skipped

    shot 13·duplicate of #12

  • 7:30 #14 done18 line(s)

    shot 14·sharpness 1087.8

    1. The solution: four ways to manage context0.99
    2. Every context engineering decision falls into one of these categories.1.00
    3. Write Context1.00
    4. Select Context1.00
    5. Persist information outside the window so1.00
    6. Pull only what's relevant for this step - not0.99
    7. it's available without filling it up1.00
    8. everything that might be useful0.98
    9. Scratchpads, memory files, CLAUDE.md0.99
    10. RAG, @-mentions, file references1.00
    11. Compress Context1.00
    12. Isolate Context1.00
    13. Reduce what's already in the window to just1.00
    14. Split work across separate sessions so no0.99
    15. the tokens the task actually needs1.00
    16. single window accumulates everything1.00
    17. Summarization, trimming logs1.00
    18. Parallel agents, task separation1.00

  • 7:48 #15 skipped

    shot 15·duplicate of #14

  • 8:14 #16 skipped

    shot 16·duplicate of #14

  • 8:44 #17 skipped

    shot 17·duplicate of #14

  • 9:17 #18 skipped

    shot 18·duplicate of #14

  • 9:26 #19 skipped

    shot 19·duplicate of #14

  • 9:51 #20 skipped

    shot 20·duplicate of #14

  • 10:14 #21 skipped

    shot 21·duplicate of #14

  • 11:10 #22 done19 line(s)

    shot 22·sharpness 1553.1

    1. The habit that ties it together0.99
    2. You don't need to think about all four strategies on every task. You need one default0.99
    3. behavior.1.00
    4. One task per session1.00
    5. Watch the meter1.00
    6. When in doubt, restart0.99
    7. Finish the task, close the0.99
    8. If you're past 50%, you're0.98
    9. Something feels off? Don't0.99
    10. session. Start fresh for1.00
    11. already in the dumb zone.0.99
    12. debug the AI. Start a new1.00
    13. the next one.0.99
    14. Wrap up or restart.0.98
    15. session.1.00
    16. Frequent new sessions aren't a sign of confusion - they're a sign0.99
    17. of discipline.0.99
    18. "Context engineering is the delicate art and science of filling the context wi1.00
    19. with just the right information for the next step." - Andrej Karpathy0.99

  • 11:44 #23 done3 line(s)

    shot 23·sharpness 409.3

    1. Research → Plan →0.96
    2. Implement1.00
    3. The workflow that went viral on HackerNews1.00

Transcript

285 cues· 4,964 words· 26,260 chars

  1. 0:00 Let's talk a little bit about what I mean by agentic engineering.
  2. 0:04 And let's maybe start with a question.
  3. 0:07 If I were to ask you right now, how are you using AI in your work?
  4. 0:11 Could you actually really explain it?
  5. 0:13 Not just, you know, it helps me code faster, it can write code really fast, but like the real workflow, what you hand off, what you keep, how you decide in between.
  6. 0:26 Most engineers can't, and that's a little wild to me because 90% of engineers are already using AI tools or have used them.
  7. 0:33 Maybe only half of them are using them on a regular basis, but that's a number that's definitely growing all the time.
  8. 0:39 And that's the current state.
  9. 0:41 So the question isn't whether your team is using AI.
  10. 0:44 They are.
  11. 0:45 The question is whether you're getting the most out of it or you're just kind of auto-completing your way through the day.
  12. 0:52 That gap between using AI and being able to articulate how you work with it, that's what this talk is all about.
  13. 1:01 And really, I think it represents a paradigm shift of how we think about AI.
  14. 1:07 And, you know, the history of AI and software engineering is moving very fast.
  15. 1:12 It's also very surprisingly short, right?
  16. 1:15 In the early 2020s, we got tools that could finish the lines for you.
  17. 1:19 You'd type, you know, half of a function signature and the model would guess the rest of it.
  18. 1:24 You know, kind of like autocomplete on steroids.
  19. 1:26 It's a neat trick.
  20. 1:29 And then in 2022, models started to be able to suggest entire functions, right?
  21. 1:33 You could describe what you wanted and chat with a model and maybe get a working implementation back.
  22. 1:39 And this is where GitHub Copilot first came on the scene and broke through and millions of developers started using it.
  23. 1:45 And for the first time, it was starting to seem like maybe AI wasn't a novelty.
  24. 1:49 Maybe it was generally useful.
  25. 1:52 But then in 2025, something really broke.
  26. 1:55 It's what we're living in now in 2026.
  27. 1:59 The models don't just suggest they can execute.
  28. 2:02 They can take a task and break it down and figure out which files need to be touched and make the changes and run the test themselves and then come back with an actual pull request.
  29. 2:13 And so that's not just fancy autocomplete.
  30. 2:16 It's not just a faster horse.
  31. 2:18 It's a collaborator.
  32. 2:19 It's a different way of working.
  33. 2:22 And Armand, the creator of Flask for those Python folks here, put it, I think, perfectly.
  34. 2:27 We're no longer just using machines.
  35. 2:30 We're now working with them.
  36. 2:32 And that framing, I think, captures this real shift, right?
  37. 2:36 Tools are things that you pick up and put down.
  38. 2:38 You use a hammer.
  39. 2:40 You don't work with a hammer.
  40. 2:42 But the AI coding agents we have today, they're kind of somewhere more in between.
  41. 2:46 And they're maybe a little bit more like working with another engineer, right?
  42. 2:51 Now it just happens to be an engineer who's read every stack overflow answer ever written.
  43. 2:57 And I think that needs a mental model shift.
  44. 3:00 And this is the mental model I want you to carry through the rest of this video and honestly through the rest of your, you know, next couple years of your career and working with these tools.
  45. 3:08 I do think they're still tools, but we have to think about them differently.
  46. 3:12 you kind of have to think about your AI agent as an energetic, enthusiastic, extremely well-read, often confidently wrong junior developer.
  47. 3:25 That junior developer is incredibly fast.
  48. 3:27 They don't easily get tired.
  49. 3:30 They don't have any ego about their code.
  50. 3:32 They'll happily rewrite something six times if you ask them to.

Open at this second