vidtheque.

read-only demo

Videos khVX_BUnEwU

Active Graph Agent Runtime (BabyAGI 4) — Yohei Nakajima, Untapped Capital

index_state ready data_status ok

AI Engineer· published 2026-07-22· 0:17:34· en-US· indexed 2026-08-10 23:52

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

52 shot(s).

keyframes kept every frame deduplicated

What was stored

cues
217
whisperx 217
chunks
31
from 217 cues
keyframes
48
kept of 52 captured
frames with text
48
1,016 lines read
chapters
11
from the source metadata
keyframe bytes
5.3 MB
word timings on 217 cues

Provenance

Each pipeline stage, its state and the model that produced it
stage state model started took
fetch done — 2026-08-10 23:48 1m 11s
stt done — 2026-08-10 23:49 27s
chunk done — 2026-08-10 23:49 0s
text_embed done — 2026-08-10 23:49 0s
keyframe done — 2026-08-10 23:49 1m 57s
ocr done — 2026-08-10 23:51 22s
frame_embed done — 2026-08-10 23:52 8s

Frames, and what the machine read

  • 0:02 #0 done2 line(s)

    shot 0·sharpness 449.7

    1. AlEngineer0.96
    2. World's Fair1.00

  • 0:03 #1 done2 line(s)

    shot 1·sharpness 662.0

    1. AIEngineer0.95
    2. World's Fair0.99

  • 0:10 #2 done24 line(s)

    shot 2·sharpness 2739.9

    1. LAB & PLATINUM SPONSORS0.99
    2. Amazon AGI Lab0.98
    3. ANTHROP\C1.00
    4. Google DeepMind1.00
    5. MINIMAX0.97
    6. OpenAI0.92
    7. Akamai1.00
    8. arize0.92
    9. aws1.00
    10. Braintrust bright data0.98
    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

  • 0:16 #3 done1 line(s)

    shot 3·sharpness 395.2

    1. World's Fair0.99

  • 0:45 #4 done10 line(s)

    shot 4·sharpness 1368.3

    1. AlEngineer0.99
    2. World's Fair0.97
    3. PRESENTED BY1.00
    4. Microsoft1.00
    5. let's build the simplest thing1.00
    6. that can build itself1.00
    7. @yoheinakajima1.00
    8. activegraph.ai1.00
    9. World's Fair0.95
    10. Engineering the future of Al0.99

  • 0:50 #5 done36 line(s)

    shot 5·sharpness 1219.4

    1. AlEngineer0.99
    2. World's Fair0.98
    3. United States - Past 5 years0.96
    4. Interest over time0.99
    5. ai agents1.00
    6. Search term1.00
    7. Let it be free!0.97
    8. Open-sourcing "Baby AGr, a paired down version of the "Task-Driven0.94
    9. Autonomous Agent* at 105 lines of code.0.95
    10. PRESENTED BY1.00
    11. 1001.00
    12. Google Trends1.00
    13. Three task agents (execution, creation, prioritization) work in harmony..0.90
    14. forever.0.97
    15. Microsoft1.00
    16. 801.00
    17. 国0.74
    18. Cit-tub - yoheinakajmalbabyagi0.79
    19. 601.00
    20. 6:28 AM - Apr 3. 2023 1.9M Vlews0.86
    21. Feb 28 - Mar 31, 20230.97
    22. 401.00
    23. ai agents 00.98
    24. Mar 31 - Apr 30, 20231.00
    25. 201.00
    26. ai agents 10.98
    27. 20221.00
    28. 20231.00
    29. 20241.00
    30. 20251.00
    31. 20261.00
    32. Google Trends1.00
    33. @yoheinakajima1.00
    34. activegraph.ai1.00
    35. World's Fair0.99
    36. Engineering the future of Al0.98

  • 1:01 #6 done29 line(s)

    shot 6·sharpness 3288.9

    1. AlEngineer0.98
    2. World's Fair0.99
    3. Amjad Masad @amasad · Mar 290.91
    4. Yohei prototyped a mini AGI on Replit0.97
    5. eDOrDEK·Mar 260.73
    6. ≡ Forbes0.98
    7. FORTUNE1.00
    8. EDITORS' PICK0.97
    9. Auto-GPT May Be1.00
    10. BabyAGI is taking Silicon Valley1.00
    11. The Strong AI Tool1.00
    12. by storm. Should we be scared?0.99
    13. That Surpasses0.97
    14. ChatGPT1.00
    15. PRESENTED BY1.00
    16. Bernard Marr Contributor0.96
    17. Microsoft1.00
    18. Auto-GPT and BabyAGI: How0.99
    19. Hype grows over "autonomous" Al agents0.99
    20. that loop GPT-4 outputs1.00
    21. generative Al to the masses1.00
    22. 'autonomous agents' are bringing0.98
    23. Auto-GPT, BabyAGI, and AgentGPT: How to use Al age0.98
    24. OpenAI1.00
    25. @yoheinakajima1.00
    26. activegraph.ai1.00
    27. World's Fair0.99
    28. TRACK 5• JULY 2, 20260.95
    29. Graphs1.00

  • 1:13 #7 done163 line(s)

    shot 7·sharpness 1096.3

    1. AlEngineer0.98
    2. World's Fair0.97
    3. BabyAGI0.99
    4. dee0.71
    5. cat0.93
    6. Deer0.85
    7. t1f0.61
    8. Fax0.91
    9. 20.96
    10. 200.94
    11. v30.91
    12. Date0.99
    13. Apr 230.84
    14. Apr 230.91
    15. May 230.89
    16. Jun 230.93
    17. Jul 230.90
    18. Sep 230.97
    19. Sep 240.98
    20. 0ct 240.86
    21. Feb 260.82
    22. Lines0.96
    23. 1051.00
    24. 3001.00
    25. 3201.00
    26. 3541.00
    27. 8870.99
    28. 2,2990.93
    29. 5,9620.88
    30. 1740.99
    31. 33.5060.99
    32. Files0.70
    33. 110.92
    34. 25-0.93
    35. 40+0.88
    36. 70-0.89
    37. PRESENTED BY1.00
    38. Model0.82
    39. Planning1.00
    40. per-loop0.94
    41. davinci-0030.98
    42. GPT-40.99
    43. dynamic replan0.94
    44. upfront0.99
    45. GPT-4 + 3.50.99
    46. GPT-3.50.97
    47. upfront0.98
    48. GPT-3.50.99
    49. upfront + reflect0.97
    50. GPT-3.5-16k0.96
    51. upfront +0.95
    52. reflect0.93
    53. any1.00
    54. any (litellm)0.99
    55. Smplicit (LLM)0.94
    56. any (litellm)0.97
    57. inplicit (LLM)0.93
    58. Microsoft1.00
    59. Execution0.99
    60. sequential0.99
    61. sequential1.00
    62. sequential0.99
    63. parallel0.97
    64. parallel0.98
    65. parallel0.98
    66. sequential0.97
    67. sequential0.98
    68. async + pool0.96
    69. Task deps1.00
    70. sone0.92
    71. single0.99
    72. multi0.92
    73. multi0.86
    74. multi0.95
    75. mlti0.91
    76. fn graph0.86
    77. none0.99
    78. objective tree0.96
    79. Teswination0.93
    80. sever0.60
    81. all complete0.99
    82. a110.79
    83. complete0.99
    84. all complete1.00
    85. all complete1.00
    86. all complete0.99
    87. manual0.98
    88. tool signal0.97
    89. end_tuan0.97
    90. Hemory0.91
    91. Pinecone1.00
    92. session str1.00
    93. dep chain1.00
    94. dep chain0.99
    95. enbedings0.97
    96. ndjson + summazy0.94
    97. exoc logs0.93
    98. messages[]0.96
    99. SQLite + KG0.95
    100. Tools0.93
    101. 15°0.79
    102. fn packs0.96
    103. self-creating0.99
    104. self-creating • persist0.97
    105. Extensibility0.94
    106. edit souzce0.90
    107. edit souace0.90
    108. edit souace0.90
    109. edit source0.87
    110. skill plugins0.99
    111. skill plugins0.99
    112. tn registry0.90
    113. zuntime exec()0.84
    114. zegistez + 080.93
    115. I/00.83
    116. CLI print0.99
    117. CLI print0.99
    118. CLI print1.00
    119. CLI + file0.95
    120. CLI - file0.90
    121. Flask web UI0.97
    122. Flask dashboard0.97
    123. CLI0.99
    124. multi-channel1.00
    125. Concurrency0.92
    126. threads0.97
    127. threads0.98
    128. threads0.95
    129. asyne + semaphore0.94
    130. Erzor handl ing0.94
    131. tzy/except0.97
    132. try/except0.95
    133. try/except1.00
    134. txy/except0.93
    135. try/except0.99
    136. logged1.00
    137. try/except1.00
    138. retxy + backoff +0.92
    139. repair0.81
    140. Key insight1.00
    141. LLMs can chain1.00
    142. tasks1.00
    143. tasks need1.00
    144. structure0.98
    145. plan0.99
    146. upfront0.94
    147. pazallelize0.98
    148. DAG1.00
    149. axchitecture0.96
    150. plugin1.00
    151. chat0.91
    152. reflection0.97
    153. functions as1.00
    154. atoms0.95
    155. LLM is the0.98
    156. plannex0.91
    157. sutonomous assistant0.98
    158. babyagi.wiki1.00
    159. @yoheinakajima1.00
    160. activegraph.ai1.00
    161. World's Fair0.95
    162. TRACK 5• JULY 2, 20260.95
    163. Graphs1.00

  • 1:19 #8 done18 line(s)

    shot 8·sharpness 1560.1

    1. AlEngineer0.98
    2. World'sFair1.00
    3. instagraph1.00
    4. mindgraph1.00
    5. code graph1.00
    6. log graph1.00
    7. 25 files across 9 nested folders with functions that depend on each0.96
    8. other, cleanly organized0.99
    9. upgade0.94
    10. created by Al based on auto-generated file descriptions which curently0.97
    11. PRESENTED BY1.00
    12. "this include param description for each function which 1 don't need for0.98
    13. Microsoft1.00
    14. @yoheinakajima1.00
    15. activegraph.ai1.00
    16. World'sFair1.00
    17. TRACK 5• JULY 2,20260.96
    18. Graphs1.00

  • 1:33 #9 done23 line(s)

    shot 9·sharpness 1059.0

    1. AlEngineer0.97
    2. World'sFair1.00
    3. graphrag1.00
    4. Search term0.99
    5. graph rag1.00
    6. United States1.00
    7. Past 5 years0.97
    8. Web Search0.96
    9. Interest over time0.99
    10. Urited States · Past 5 years0.94
    11. Codegraph1.00
    12. Log graph1.00
    13. Mindgraph1.00
    14. 3/21/20241.00
    15. 4/11/20241.00
    16. 3/16/20241.00
    17. Instagraph1.00
    18. 9/13/20231.00
    19. @yoheinakajima1.00
    20. activegraph.ai1.00
    21. World's Fair0.99
    22. TRACK 5· JULY 2, 20260.95
    23. Graphs1.00

  • 1:46 #10 done31 line(s)

    shot 10·sharpness 1814.6

    1. AlEngineer0.99
    2. World'sFair1.00
    3. E2B0.99
    4. Firecrawl1.00
    5. Composio1.00
    6. mem01.00
    7. FalkorDB1.00
    8. AgentMail0.95
    9. Browser Use0.96
    10. The General Intelligence0.97
    11. Company of New York0.99
    12. Hamster1.00
    13. Intelligent Internet1.00
    14. FlymyAl0.94
    15. DENIED1.00
    16. ANTIMLABs0.87
    17. Zinley1.00
    18. AGNOST AI1.00
    19. ∏∑0.75
    20. COVENANT LABS1.00
    21. SAZABI1.00
    22. VideoDB1.00
    23. Monid1.00
    24. RunAnywhere1.00
    25. untapped.vc1.00
    26. agentfund.com1.00
    27. @yoheinakajima1.00
    28. activegraph.ai1.00
    29. World's Fair0.98
    30. TRACK 5• JULY 2, 20260.95
    31. Graphs1.00

  • 1:57 #11 done10 line(s)

    shot 11·sharpness 943.0

    1. AlEngineer0.98
    2. World'sFair0.97
    3. ActiveGraph1.00
    4. an event-sourced graph runtime for1.00
    5. building auditable agents1.00
    6. @yoheinakajima1.00
    7. activegraph.ai1.00
    8. World's Fair0.98
    9. TRACK 5• JULY 2, 20260.95
    10. Graphs1.00

  • 2:03 #12 done19 line(s)

    shot 12·sharpness 930.6

    1. AlEngineer1.00
    2. World's Fair0.98
    3. The Log is the Agent'0.98
    4. Event-Sourced Reactive Graphs for Auditable, Forkable Agentie Systems0.94
    5. Yobei Nakajima0.87
    6. Untappod Capital0.95
    7. activegraph..i0.95
    8. 2199 11[1] 102660.64
    9. May 22, 20260.99
    10. Abetraet0.80
    11. an1.00
    12. hility, with state persiso0.86
    13. for0.94
    14. s, LLMI-barkond routinns, or logic0.87
    15. @yoheinakajima1.00
    16. activegraph.ai1.00
    17. World's Fair0.97
    18. TRACK 5· JULY 2, 20260.95
    19. Graphs1.00

  • 2:17 #13 done13 line(s)

    shot 13·sharpness 1312.8

    1. AlEngineer0.98
    2. World'sFair1.00
    3. today, most people build agents around the llm0.99
    4. Logs1.00
    5. Response APl0.97
    6. Memory1.00
    7. LLM1.00
    8. Tools1.00
    9. @yoheinakajima1.00
    10. activegraph.ai1.00
    11. World'sFair1.00
    12. TRACK 5• JULY 2, 20260.95
    13. Graphs1.00

  • 2:53 #14 done18 line(s)

    shot 14·sharpness 1355.9

    1. AlEngineer0.99
    2. World's Fair0.96
    3. this projects a graph (state of agent)0.99
    4. agent state1.00
    5. as a graph0.99
    6. Logs1.00
    7. goal.created1.00
    8. prompt.created1.00
    9. Ilm.called0.97
    10. tool.called1.00
    11. tool.response1.00
    12. Ilm.response0.96
    13. prompt.updated1.00
    14. @yoheinakajima1.00
    15. activegraph.ai1.00
    16. World's Fair0.96
    17. TRACK 5• JULY 2, 20260.97
    18. Graphs1.00

  • 3:14 #15 done20 line(s)

    shot 15·sharpness 1794.6

    1. AlEngineer0.99
    2. World's Fair0.98
    3. attach behaviors that react to graph changes0.98
    4. behaviors can then emit events1.00
    5. agent state1.00
    6. as a graph0.99
    7. behavior1.00
    8. Logs1.00
    9. goal.created1.00
    10. prompt.created1.00
    11. Ilm.called0.97
    12. tool.called1.00
    13. tool.response1.00
    14. Ilm.response0.95
    15. prompt.updated1.00
    16. @yoheinakajima1.00
    17. activegraph.ai1.00
    18. World's Fair0.96
    19. TRACK 5• JULY 2, 20260.95
    20. Graphs1.00

  • 3:45 #16 done25 line(s)

    shot 16·sharpness 1763.2

    1. AlEngineer0.98
    2. World'sFair1.00
    3. policies determine how the graph can be modified0.98
    4. agent state0.98
    5. as a graph0.99
    6. behavior1.00
    7. behavior1.00
    8. behavior1.00
    9. behavior1.00
    10. behavior1.00
    11. behavior1.00
    12. policies1.00
    13. Logs1.00
    14. goal.created1.00
    15. prompt.created1.00
    16. Ilm.called0.97
    17. tool.called1.00
    18. tool.response1.00
    19. Ilm.response0.96
    20. prompt.updated1.00
    21. @yoheinakajima1.00
    22. activegraph.ai1.00
    23. World's Fair0.97
    24. TRACK 5• JULY 2, 20260.94
    25. Graphs1.00

  • 4:11 #17 done28 line(s)

    shot 17·sharpness 2003.2

    1. AlEngineer0.99
    2. World'sFair1.00
    3. this is your agent on activegraph0.99
    4. behavior1.00
    5. behavior1.00
    6. behavior1.00
    7. behavior1.00
    8. behavior1.00
    9. behavior1.00
    10. policies1.00
    11. Logs1.00
    12. goal.created1.00
    13. prompt.created1.00
    14. Ilm.called0.95
    15. tool.called0.98
    16. tool.response1.00
    17. Ilm.response0.96
    18. prompt.updated1.00
    19. tool.called1.00
    20. tool.response1.00
    21. Ilm.response0.98
    22. prompt.updated1.00
    23. you get replays, rollbacks, and forks0.99
    24. @yoheinakajima1.00
    25. activegraph.ai1.00
    26. World's Fair0.96
    27. TRACK 5• JULY 2, 20260.96
    28. Graphs0.93

  • 4:18 #18 done8 line(s)

    shot 18·sharpness 838.2

    1. AlEngineer0.97
    2. World'sFair1.00
    3. Logs1.00
    4. @yoheinakajima1.00
    5. activegraph.ai1.00
    6. World'sFair1.00
    7. TRACK 5· JULY 2, 20260.95
    8. Graphs1.00

  • 4:42 #19 done39 line(s)

    shot 19·sharpness 959.2

    1. AlEngineer0.99
    2. World's Fair0.97
    3. replace messages[ ] with logs0.96
    4. run - "Evaluate Acme"0.93
    5. Reset1.00
    6. PLay0.92
    7. Step -0.89
    8. NAIVE · MESSAGES[)0.88
    9. EVENT LOG - SOURCE OF TRUTH0.96
    10. USER Evaluate Acme as an investment.0.98
    11. goal.created0.96
    12. goal:"Évaluate Acne"0.92
    13. ASSISTANT Plan: (1) research market (2) draft memo0.97
    14. behavior.started0.99
    15. ASSISTANT Market LoOks -$4B TAM.0.92
    16. planser0.95
    17. To0. fetched q3_report.pdf (12kb)0.92
    18. object.created0.96
    19. task research0.99
    20. AsSIsTant _(earlier turns dropped to fit windaw)_0.90
    21. abject.created0.99
    22. ASSISTANT Final memo drafted.0.99
    23. task • mno -depends_on0.85
    24. SYSTER run complete0.94
    25. lle.responded0.91
    26. claim: TAM -8480.84
    27. tool.responded0.98
    28. g3_report.pdf0.93
    29. object.created0.98
    30. clain • incumbest risk0.91
    31. patch.applied1.00
    32. meno - status=done0.83
    33. runtime.idle0.95
    34. no pending events0.97
    35. @yoheinakajima1.00
    36. activegraph.ai1.00
    37. World's Fair0.97
    38. TRACK 5• JULY 2, 20260.95
    39. Graphs1.00

  • 4:59 #20 done17 line(s)

    shot 20·sharpness 1456.9

    1. AlEngineer0.97
    2. World'sFair1.00
    3. you can't edit the graph, just emit events0.99
    4. # you never set fields directly - you emit events, the graph folds them0.98
    5. research = graph.add_object("task", {"title": "Research", "status": "open"})0.99
    6. meno0.99
    7. = graph.add_object("task", {"title": "Draft memo", "status": "blocked"})0.99
    8. graph.add_relation(research.id, memo.id, "depends_on")1.00
    9. # reads are queries over the projected world1.00
    10. open_tasks = graph.objects(type="task", where={"status": "open"})0.99
    11. blockers = graph.relations(type="depends_on", target=memo.id)0.99
    12. but you query over the graph1.00
    13. @yoheinakajima1.00
    14. activegraph.ai1.00
    15. World'sFair1.00
    16. TRACK 5· JULY 2,20260.95
    17. Graphs1.00

  • 5:25 #21 done21 line(s)

    shot 21·sharpness 1866.1

    1. AlEngineer0.97
    2. World'sFair1.00
    3. behaviors listen to graph changes and emit events0.99
    4. @behavior(name="planner", on=["goal.created"])0.98
    5. def planner(event, graph, ctx):1.00
    6. research = graph.add_object("task", ("title": "Research", "status": "open"})0.97
    7. PRESENTED BY1.00
    8. meno0.94
    9. = graph.add_object("task", {("title": "Draft nemo", "status": "blocked"})0.95
    10. Microsoft1.00
    11. graph.add_relation(research.id, memo.id, "depends_on")0.99
    12. ends_on", on=["task.completed"])0.98
    13. def unblock(relation, event, graph, ctx):0.99
    14. if event.payload["task_id"] = relation.source:0.99
    15. graph.patch_object(relation.target, ("status": "open"})0.98
    16. behaviors can live on edges as a relation_behavior0.99
    17. @yoheinakajima1.00
    18. activegraph.ai1.00
    19. World'sFair1.00
    20. TRACK 5• JULY 2,20260.96
    21. Graphs1.00

  • 5:34 #22 done16 line(s)

    shot 22·sharpness 1308.2

    1. AlEngineer0.98
    2. World'sFair1.00
    3. behavior subscriptions can be complex1.00
    4. from activegraph import behavior0.99
    5. @behavior(1.00
    6. PRESENTED BY1.00
    7. name="contradiction_detector",1.00
    8. on=["object.created"],1.00
    9. Microsoft1.00
    10. where={"object.type": "claim"},1.00
    11. pattern="(c:claim)-[:contradicts]->(other:claim)",1.00
    12. @yoheinakajima1.00
    13. activegraph.ai1.00
    14. World'sFair1.00
    15. TRACK 5• JULY 2,20260.95
    16. Graphs1.00

  • 5:46 #23 skipped

    shot 23·duplicate of #22

Transcript

217 cues· 3,410 words· 18,380 chars

  1. 0:12 Hi, everybody, thanks for coming.
  2. 0:14 I'm excited to be here.
  3. 0:16 AI Engineer World Fair has been so fun meeting everybody.
  4. 0:20 But I'm here to talk about Active Graph, which is my new open source experimental approach to building agents, which looks a little bit different than maybe you've been building agents.
  5. 0:29 It's definitely experimental.
  6. 0:31 The idea is more to inspire you with some potentially new ideas.
  7. 0:36 Agents are awesome, but long-running agents break.
  8. 0:39 And if they're so awesome, why am I still building them?
  9. 0:42 They should build themselves.
  10. 0:44 Let's build the simplest thing that can build itself has basically been kind of my research theme for the last three years since I did Baby AGI back in March of 2023.
  11. 0:53 So that's over three years ago.
  12. 0:55 If you were there at the time, it was crazy.
  13. 0:57 It went wild.
  14. 0:58 It was covered by media.
  15. 1:00 People thought it was going to work.
  16. 1:01 It didn't work at all.
  17. 1:04 Over the course of three years, I've done nine iterations of Baby AGI, less fanfare, but every time just experimenting on how do we get autonomous agents to actually work, usually with the theme of self-improvement.
  18. 1:14 If you go to Baby AGI Wiki, you can see earlier experiments.
  19. 1:18 In this process, I kept coming back to graphs, and I've had a couple of projects.
  20. 1:23 Earlier, I did one called Instagraph and MineGraph that was pre-GraphRagRag.
  21. 1:27 I did some code graphs, function graphs, log graphs, and since then it seems like a lot of people have started using graphs to build agents.
  22. 1:36 And in addition to that, I've actually gotten to invest in a good number of agentic companies, some of which I'm sure you recognize through my fund's untapped capital, and I also have an agent fund.
  23. 1:47 But yeah, that's me.
  24. 1:48 Yohei, VC by day, Builder by night.
  25. 1:51 You might recognize this face more than this face.
  26. 1:55 Active Graph is an event-sourced graph runtime for building auditable data.
  27. 2:00 I have a paper that was my first archive paper called the log is the agent, but I'm here to explain it.
  28. 2:06 So today most people build agents around the LLM.
  29. 2:08 You start with the LLM, you add a response API, you give it tools, you add memory, and then you make sure you log everything correctly, which can give you all the benefits the Active Graph will give you.
  30. 2:19 But Active Graph asks what if you built around the log?
  31. 2:22 Now, what does that mean?
  32. 2:24 It means not everything the agent does, but more importantly, every change to the agent, right?
  33. 2:31 Nobody here is using the same agent they were using a year ago.
  34. 2:34 And the agent you're gonna use a year from now is gonna be different.
  35. 2:37 And a lot of people, what the agent does and how the agent changes are tracked in two different places.
  36. 2:41 But I'm saying let's flatten that down into a single immutable event log.
  37. 2:46 And this is the ground truth of the agent.
  38. 2:49 And this projects a sort of graph.
  39. 2:50 This is the state of the agent.
  40. 2:52 And what I mean by that is, for example, a prompt can be edited multiple times, but you might have a master prompt that gets used when you query the graph.
  41. 3:00 And then on top of this, you attach something that I'm calling behaviors.
  42. 3:04 Behaviors react to graph changes.
  43. 3:09 And then they emit events, which then in turn updates the state of the agent, which might trigger new behaviors.
  44. 3:18 LLMs don't talk to each other in Active Graph.
  45. 3:20 They all communicate through this shared state, and that's what makes it a little bit different.
  46. 3:24 Behaviors can be deterministic, or they can include LLMs, which is how you build this agent.
  47. 3:30 And you get this beautiful typed
  48. 3:33 Event log, that's the source of truth about everything the agent did and everything, every change that's happened, which means, actually, oh shoot, I jumped ahead.
  49. 3:41 So in addition to that, there's a concept called policies which determine how the graph can be modified.
  50. 3:45 I'll come back to it, but for example, things like a source article that you found in research, you might be fine with adding, but if you're changing a prompt, maybe you want human in the loop,

Chapters

  1. 0:00 ActiveGraph and three years of BabyAGI
  2. 1:55 Build around the log, not the LLM
  3. 3:24 Behaviors, policies, and views
  4. 6:43 Packs and the blackboard architecture lineage
  5. 8:12 The log as memory, and the API key that resumed itself
  6. 9:53 Reference agents built natively on the log
  7. 11:11 Self improvement: regimes and controlled self modification
  8. 12:25 ActiveGraph Lab writes its own experiments
  9. 13:02 A Pokemon card competition as a testbed
  10. 14:33 Surprises: why AI architects this better
  11. 15:49 Why an agent needs an experiential world model

Open at this second