Videos WLXxTaPagA8

Every Solo Agent Builder Eventually Reinvents a Worse Version of CI/CD - Sumaiya Shrabony

index_state ready data_status ok

AI Engineer· published 2026-07-11· 0:10:51· en-US· indexed 2026-08-11 04:09

Open on YouTube

85 shot(s).

keyframes kept every frame deduplicated

What was stored

cues: 161; whisperx 161
chunks: 20; from 161 cues
keyframes: 78; kept of 85 captured
frames with text: 47; 902 lines read
chapters: 0; from the source metadata
keyframe bytes: 12.6 MB; word timings on 161 cues

Provenance

Each pipeline stage, its state and the model that produced it
stage	state	model	started	took
`fetch`	done	—	2026-08-11 04:05	1m 47s
`stt`	done	—	2026-08-11 04:06	11s
`chunk`	done	—	2026-08-11 04:07	0s
`text_embed`	done	—	2026-08-11 04:07	0s
`keyframe`	done	—	2026-08-11 04:07	1m 07s
`ocr`	done	—	2026-08-11 04:08	39s
`frame_embed`	done	—	2026-08-11 04:08	12s

Frames, and what the machine read

0:14 #0 empty

shot 0·sharpness 341.1
0:20 #1 done10 line(s)

shot 1·sharpness 1277.9
1. AI ENGINEER WORLD'S FAIR 20260.99
2. Every Solo Agent Builder1.00
3. Eventually Reinvents a Worse1.00
4. Version of Cl/CD0.97
5. Agent demos end at the happy path. Agent operations start when the0.99
6. happy path lies to you.0.99
7. SUMAIYA SHRABONY1.00
8. Technical Program Manager ·University of Colorado Denver0.99
9. 19-skill open-source agent system·Ground Truth newsletter1.00
10. github.com/safrin96/agentic-content-system1.00
0:21 #2 done10 line(s)

shot 2·sharpness 1121.9
1. AI ENGINEER WORLD'S FAIR 20260.99
2. Every Solo Agent Builder0.99
3. Eventually Reinvents a Worse1.00
4. Version of Cl/CD0.97
5. Agent demos end at the happy path. Agent operations start when the0.99
6. happy path lies to you.0.99
7. SUMAIYA SHRABONY1.00
8. Technical Program Manager · University of Colorado Denver0.98
9. 19-skill open-source agent system· Ground Truth newsletter0.99
10. github.com/safrin96/agentic-content-system1.00
0:25 #3 done10 line(s)

shot 3·sharpness 1001.6
1. AI ENGINEER WORLD'S FAIR 20260.99
2. Every Solo Agent Builder1.00
3. Eventually Reinvents a Worse1.00
4. Version of Cl/CD0.97
5. Agent demos end at the happy path. Agent operations start when the0.99
6. happy path lies to you.0.99
7. SUMAIYA SHRABONY1.00
8. Technical Program Manager · University of Colorado Denver0.99
9. 19-skill open-source agent system· Ground Truth newsletter0.98
10. github.com/safrin96/agentic-content-system1.00
0:30 #4 skipped

shot 4·duplicate of #0
0:41 #5 empty

shot 5·sharpness 315.2
0:48 #6 empty

shot 6·sharpness 341.0
0:55 #7 empty

shot 7·sharpness 358.2
0:58 #8 empty

shot 8·sharpness 370.8
1:03 #9 empty

shot 9·sharpness 370.4
1:07 #10 empty

shot 10·sharpness 367.2
1:11 #11 empty

shot 11·sharpness 391.6
1:14 #12 skipped

shot 12·duplicate of #11
1:22 #13 empty

shot 13·sharpness 315.7
1:28 #14 done22 line(s)

shot 14·sharpness 1325.3
1. THE SYSTEM1.00
2. A production workflow with seven0.99
3. handoffs1.00
4. Every handoff is a place where the system can silently lie to you.0.99
5. CONTRACT BREAK1.00
6. MISSED VIOLATION0.99
7. FALSE PASS1.00
8. Scheduler1.00
9. Command1.00
10. Research1.00
11. →1.00
12. Plan1.00
13. Production Skill1.00
14. →0.99
15. Verifier1.00
16. Reviewer1.00
17. →1.00
18. Output1.00
19. This system runs every other Saturday. Reads from a knowledge vault. Creates research briefs.0.99
20. Builds a content plan. Produces 12 pieces. Runs verifier passes, reviewer gates, deduplication.1.00
21. Saves structured markdown outputs.1.00
22. The content is not the point. The handoffs are the point.1.00
1:39 #15 done26 line(s)

shot 15·sharpness 1399.5
1. THE SYSTEM0.99
2. A production workflow with seven1.00
3. handoffs1.00
4. Every handoff is a place where the system can silently lie toyou.0.97
5. CONTRACT BREAK0.98
6. MISSED VIOLATION0.99
7. FALSE PASS0.95
8. Scheduler1.00
9. →1.00
10. Command1.00
11. →1.00
12. Research1.00
13. →1.00
14. Plan1.00
15. →0.99
16. Production Skill1.00
17. →0.99
18. Verifier1.00
19. →0.99
20. Reviewer1.00
21. →1.00
22. Output1.00
23. This system runs every other Saturday. Reads from a knowledge vault. Creates research briefs.0.99
24. Builds a content plan. Produces 12 pieces. Runs verifier passes, reviewer gates, deduplication.1.00
25. Saves structured markdown outputs.1.00
26. The content is not the point. The handoffs are the point.0.99
1:41 #16 empty

shot 16·sharpness 362.4
1:49 #17 done25 line(s)

shot 17·sharpness 2020.5
1. THE SYSTEM1.00
2. A production workflow with seven0.99
3. handoffs1.00
4. Every handoff is a place where the system can silently lie to you.0.99
5. CONTRACT BREAK0.99
6. MISSED VIOLATION1.00
7. FALSE PASS1.00
8. Scheduler1.00
9. →1.00
10. Command1.00
11. →1.00
12. Research1.00
13. →1.00
14. Plan1.00
15. Production Skill0.98
16. →0.99
17. Verifier1.00
18. →0.99
19. Reviewer0.95
20. →1.00
21. Output1.00
22. This system runs every other Saturday. Reads from a knowledge vault. Creates research briefs.0.99
23. Builds a content plan. Produces 12 pieces. Runs verifier passes, reviewer gates, deduplication.0.99
24. Saves structured markdown outputs.1.00
25. The content is not the point. The handoffs are the point.0.99
1:52 #18 empty

shot 18·sharpness 354.3
1:55 #19 empty

shot 19·sharpness 337.7
1:57 #20 done20 line(s)

shot 20·sharpness 329.5
1. THE FIVE REINVENTIONS0.98
2. What you'll accidentally rebuild1.00
3. In roughly this order. Every time. Click a card to see the story or run1.00
4. the auto-tour.1.00
5. Auto-tour1.00
6. SPEED1.00
7. 1×0.99
8. Both1.00
9. Agent only0.99
10. Highlight Cl/CD0.96
11. 011.00
12. Regression Checks1.00
13. Scheduled Monitoring1.00
14. You change a prompt. Something downstream breaks. You1.00
15. Your cron job silently falls. You don't notice for a week. You0.97
16. build a way to test if the output still matches the expected0.99
17. build alerts.1.00
18. shape.1.00
19. ≈CI Monitoring0.96
20. ≈Regression Testing0.98
2:00 #21 done23 line(s)

shot 21·sharpness 871.5
1. THE FIVE REINVENTIONS1.00
2. What you'll accidentally rebuild0.99
3. In roughly this order. Every time. Click a card to see the story or run1.00
4. the auto-tour.1.00
5. Auto-tour1.00
6. SPEED1.00
7. 1×0.90
8. Both1.00
9. Agent only1.00
10. Highlight Cl/CD0.98
11. 011.00
12. 021.00
13. Regression Checks1.00
14. Scheduled Monitoring1.00
15. You change a prompt. Something downstream breaks. You1.00
16. Your cron job silently fails. You don't notice for a week. You1.00
17. build a way to test if the output still matches the expected0.99
18. build alerts.0.97
19. shape.1.00
20. ≈ CI Monitoring0.96
21. ≈ Regression Testing1.00
22. SHOW THE STORY1.00
23. +0.95
2:01 #22 done24 line(s)

shot 22·sharpness 773.1
1. Wliat yuu Il acciucillally Tcvullu0.71
2. In roughly this order. Every time. Click a card to see the story or run1.00
3. the auto-tour.1.00
4. Auto-tour1.00
5. SPEED1.00
6. 1x0.78
7. Both1.00
8. Agent only0.99
9. Highlight Cl/CD0.96
10. 011.00
11. 021.00
12. Regression Checks1.00
13. Scheduled Monitoring0.98
14. You change a prompt. Something downstream breaks. You0.99
15. Your cron job silently fails. You don't notice for a week. You0.99
16. build a way to test if the output still matches the expected0.99
17. build alerts.1.00
18. shape.1.00
19. ≈ CI Monitoring0.96
20. ≈ Regression Testing0.99
21. SHOW THE STORY1.00
22. +0.95
23. SHOW THE STORY1.00
24. +0.95
2:12 #23 done24 line(s)

shot 23·sharpness 778.2
1. 011.00
2. Regression Checks1.00
3. Scheduled Monitoring1.00
4. You change a prompt. Something downstream breaks. You0.99
5. Your cron job silently fails. You don't notice for a week. You0.98
6. build a way to test if the output still matches the expected1.00
7. build alerts.1.00
8. shape.1.00
9. = Cl Monitoring0.92
10. ≈ Regression Testing0.99
11. SHOW THE STORY0.99
12. SHOW THE STORY1.00
13. x0.93
14. YOU NOTICE0.99
15. A prompt tweak quietly mangles the1.00
16. A0.88
17. markdown structure two skills downstream.0.99
18. YOU BUILD1.00
19. A small harness that re-runs sample inputs0.99
20. and diffs the output shape against last1.00
21. week's.1.00
22. ALREADY EXISTED AS0.99
23. A regression test suite. With fixtures. Since0.99
24. pytest was a thing.0.99

Transcript

161 cues· 1,508 words· 8,690 chars

chunk 0 · 0:00–0:44 · 94 words · 603 chars
0:00 Here's something nobody warns you about when you start building agents alone.
0:04 You think you're building prompts.
0:06 You think you're building skills or workflow.
0:10 If you build long enough, especially alone,
0:13 you will start building something completely different, something that looks suspiciously like CICD, except worse, because you're building it from scratch, one failure at a time.
0:27 I'm Sumaya.
0:27 I run a 19-scale cloud code agent system.
chunk 1 · 0:31–1:14 · 102 words · 602 chars
0:31 Writing, Research, Vault Sync, Analytics Sync, Hook, Transcript, and many more.
0:39 The most useful thing I learned from building this was not how to build better prompts.
0:45 It was recognizing the five controls I was rebuilding badly and what you can do instead.
0:51 Before I show you the problem, let me show you the system that taught me the problem.
0:56 This is the agent content system.
0:59 It is open source.
1:00 Link in the description.
chunk 2 · 1:02–1:44 · 98 words · 614 chars
1:02 It runs every other siren.
1:04 It reads from a knowledge vault, creates a research brief, builds content plan, produces 12 content placements, then runs verifier passes, reviewer gates,
1:15 deduplication, and finally saves the output as markdown files.
1:19 But here's the thing that matters for this talk, not the content.
1:23 What matters is that this system has seven handoffs.
1:27 Scheduler to command, command to research, research to content plan, content plan to production skill, production skill to verifier, verifier to reviewer, reviewer to the output pool.
chunk 3 · 1:40–2:21 · 103 words · 581 chars
1:40 Every single handoff is the place where the system can lie to you.
1:45 And if you're building the system alone, nobody catches the lies except you, usually after the damage is done.
1:52 Here's the pattern I want you to watch for in your own system.
1:56 If you build agents independently, you will rebuild these five things roughly in this order.
2:04 You change a prompt or a skill and something downstream breaks.
2:09 So you build a way to test whether the output still matches the expected stream.
chunk 4 · 2:14–2:55 · 91 words · 539 chars
2:14 Congratulations, you have reinvented regression testing.
2:19 So you set up a cron job or a scheduled task.
2:22 One day it silently fails, but you haven't noticed for a week.
2:27 So you build alerts.
2:29 You just reinvented CI monitoring.
2:32 one skill changes its output schema so three skills downstream break you decided to add a validation at the boundary because of it
2:42 You just reinvented contract testing.
chunk 5 · 2:45–3:24 · 89 words · 534 chars
2:45 An artifact looks done, but it shouldn't ship.
2:48 So you add a checkpoint before it goes to the ready folder.
2:53 You just reinvented staging environments.
2:55 Something goes wrong, but you cannot find out which prompt, which skill, or which handoff had the bad output.
3:04 So you start logging everything.
3:07 she just reinvented audit rails.
3:09 The reason the title says ORS version isn't because agents are software builds.
chunk 6 · 3:15–3:59 · 98 words · 551 chars
3:15 It's because you end up needing the exact same operational guarantees.
3:20 However, the agent systems give you none of it by default.
3:25 So you build them independently, the worst version without even realizing that you are building it.
3:33 The dangerous failure in an agent system is never a bad output.
3:39 A bad output is very easy to fix.
3:42 You glance at it and immediately you can understand it's a bad output.
chunk 7 · 3:46–4:28 · 110 words · 585 chars
3:46 The dangerous failure is a polished artifact that looks great at a glance.
3:53 However,
3:54 It will never pass your exit gates.
3:57 It uses the wrong voice pattern.

Every Solo Agent Builder Eventually Reinvents a Worse Version of CI/CD - Sumaiya Shrabony

Scene timeline

What was stored

Provenance

Frames, and what the machine read

Transcript