Videos OqM67QG_Ikk

From fork() to Fleet: Designing an Agent Sandbox Cloud — Abhishek Bhardwaj, OpenAI

index_state ready data_status ok

AI Engineer· published 2026-07-13· 0:44:33· en-US· indexed 2026-08-10 19:48

Open on YouTube

95 shot(s).

keyframes kept every frame deduplicated

What was stored

cues: 520; whisperx 520
chunks: 80; from 520 cues
keyframes: 53; kept of 95 captured
frames with text: 53; 1,439 lines read
chapters: 21; from the source metadata
keyframe bytes: 11.1 MB; word timings on 520 cues

Provenance

Each pipeline stage, its state and the model that produced it
stage	state	model	started	took
`fetch`	done	—	2026-08-10 11:27	2m 08s
`stt`	done	—	2026-08-10 11:29	49s
`chunk`	done	—	2026-08-10 11:30	0s
`text_embed`	done	—	2026-08-10 19:48	1s
`keyframe`	done	—	2026-08-10 11:30	4m 04s
`ocr`	done	—	2026-08-10 11:34	24s
`frame_embed`	done	—	2026-08-10 19:48	9s

Frames, and what the machine read

0:02 #0 done2 line(s)

shot 0·sharpness 451.4
1. AlEngineer0.95
2. World's Fair0.98
0:03 #1 done2 line(s)

shot 1·sharpness 657.9
1. AlEngineer0.96
2. World's Fair0.99
0:10 #2 done24 line(s)

shot 2·sharpness 2727.9
1. LAB & PLATINUM SPONSORS0.98
2. Amazon AGI Lab0.98
3. ANTHROP\C1.00
4. Google DeepMind1.00
5. MINIMAX0.97
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
0:14 #3 done2 line(s)

shot 3·sharpness 151.0
1. AlEngineer1.00
2. World's Fair0.99
0:17 #4 done2 line(s)

shot 4·sharpness 181.4
1. AlEngineer0.97
2. World's Fair1.00
1:09 #5 done14 line(s)

shot 5·sharpness 2415.7
1. OpenAI0.93
2. AlEngineer0.98
3. World'sFair1.00
4. From fork() to fleet:1.00
5. designing an agent1.00
6. sandbox cloud1.00
7. PRESENTED·BY0.97
8. Microsoft1.00
9. Abhishek Bhardwaj1.00
10. Member of Technical Staff,1.00
11. Reinforcement Learning and Agent Infrastructure1.00
12. TRACK 1· JULY 1, 20260.94
13. World's Fair0.96
14. Sandbox & Platform Engineering0.99
1:20 #6 done9 line(s)

shot 6·sharpness 1894.7
1. AlEngineer0.99
2. Last years' AIE talk0.98
3. World'sFair1.00
4. Arrakis: how to build an Al sandbox from scratch1.00
5. PRESENTEDBY1.00
6. Microsoft1.00
7. TRACK 1· JULY 1, 20260.95
8. World'sFair1.00
9. Sandbox & Platform Engineering1.00
1:43 #7 done32 line(s)

shot 7·sharpness 2770.5
1. AlEngineer0.99
2. WHYSANDBOXES·TRAINING1.00
3. World'sFair1.00
4. Why the explosion of agent sandboxes?1.00
5. Reasoning models can emit a tool and its parameters. A harness executes the call;0.99
6. a grader scores the result; training adjusts the weights.1.00
7. PRESENTEDBY1.00
8. Update weights1.00
9. Microsoft1.00
10. RL framework0.96
11. Task1.00
12. Harness1.00
13. Call model1.00
14. Model1.00
15. training loop1.00
16. calls model and tools1.00
17. Response / tool call0.98
18. emits tool +1.00
19. parameters1.00
20. Reward1.00
21. Result1.00
22. Execute tool call1.00
23. Observation1.00
24. Grader1.00
25. Tool runtime1.00
26. grades result1.00
27. sandbox executes the call1.00
28. Where and how does this run?0.99
29. OpenAl0.96
30. TRACK 1· JULY 1,20260.95
31. World's Fair0.96
32. Sandbox & Platform Engineering0.99
2:08 #8 skipped

shot 8·duplicate of #7
2:36 #9 skipped

shot 9·duplicate of #7
3:17 #10 skipped

shot 10·duplicate of #7
3:45 #11 skipped

shot 11·duplicate of #7
4:12 #12 skipped

shot 12·duplicate of #7
4:30 #13 skipped

shot 13·duplicate of #7
5:08 #14 done29 line(s)

shot 14·sharpness 2496.6
1. AlEngineer0.99
2. WHY SANDBOXES·INFERENCE0.98
3. Why the explosion of agent sandboxes?1.00
4. World's Fair0.99
5. At inference time, tool calls are still model output. The harness executes0.99
6. them on the model's behalf and returns the observation.1.00
7. PRESENTED·BY0.95
8. Microsoft1.00
9. Task1.00
10. Task1.00
11. Harness1.00
12. Call model1.00
13. Model1.00
14. user request + context0.97
15. Result1.00
16. calls model and tools1.00
17. Response / tool call0.98
18. returns answer or0.98
19. tool call1.00
20. Parse and execute1.00
21. tool call_0.91
22. Observation1.00
23. Tool runtime1.00
24. sandbox executes the call1.00
25. Where and how does this run?1.00
26. OpenAl0.97
27. TRACK 1· JULY 1,20260.96
28. World's Fair0.99
29. Sandbox & Platform Engineering1.00
5:33 #15 skipped

shot 15·duplicate of #14
6:04 #16 done18 line(s)

shot 16·sharpness 2051.3
1. AlEngineer0.99
2. AGENT CLOUDS · LOCAL TO CLOUD0.98
3. World'sFair1.00
4. OpenClaw:1.00
5. a peek into agent clouds1.00
6. OpenClaws1.00
7. Keep your agents1.00
8. running 24/7.1.00
9. A precision claw that keeps your Mac0.98
10. slightly open so background agents can stay alive.1.00
11. PRESENTED BY0.96
12. Only $1990.89
13. Microsoft1.00
14. Pre-order now0.99
15. OpenAl0.96
16. TRACK 1• JULY 1, 20260.97
17. World's Fair0.96
18. Sandbox & Platform Engineering0.98
6:39 #17 skipped

shot 17·duplicate of #16
7:01 #18 done32 line(s)

shot 18·sharpness 2013.1
1. AGENT CLOUDS · LOCAL TO CLOUD0.95
2. AlEngineer0.99
3. Research vs. product sandbox platforms1.00
4. World's Fair0.99
5. Research1.00
6. Product1.00
7. PRESENTEDBY1.00
8. Primary optimization1.00
9. Throughput1.00
10. Latency1.00
11. Microsoft1.00
12. Why1.00
13. Many parallel rollouts1.00
14. People see results fast0.99
15. Keep GPU compute utilized1.00
16. Responsive product experience1.00
17. Reliability1.00
18. Avoid wasted tokens1.00
19. Paramount1.00
20. Avoid restarting rollouts1.00
21. Unreliable products fail1.00
22. Prevent generated code from gaining root0.98
23. Protect the host and platform resources0.99
24. Security1.00
25. Protect internal resources1.00
26. Isolate other production resources0.99
27. Resist reward hacking0.99
28. Protect user data1.00
29. OpenAl0.96
30. TRACK 1· JULY 1, 20260.95
31. World's Fair0.99
32. Sandbox & Platform Engineering0.99
7:28 #19 skipped

shot 19·duplicate of #18
7:53 #20 skipped

shot 20·duplicate of #18
8:11 #21 skipped

shot 21·duplicate of #18
8:28 #22 done27 line(s)

shot 22·sharpness 2296.0
1. AlEngineer1.00
2. AGENT PLATFORM1.00
3. Runtime, persistence and orchestration1.00
4. World's Fair0.99
5. AGENT SANDBOX CLOUD0.98
6. PRESENTED·BY0.96
7. Microsoft1.00
8. RUNTIME1.00
9. PERSISTENCE1.00
10. ORCHESTRATION1.00
11. HOW SANDBOXES RUN0.99
12. HOW AGENT STATE SURVIVES1.00
13. WHERE SANDBOXES RUN0.99
14. SECURE EXECUTION0.98
15. SAVE SANDBOX STATE0.99
16. CLUSTER SELECTION1.00
17. LOW-LATENCY CREATION1.00
18. LONG-RUNNING TASKS1.00
19. NODE SELECTION1.00
20. LOW-LATENCY TOOL EXECUTION1.00
21. RELIABLE RECOVERY1.00
22. LOW-LATENCY CREATES1.00
23. SECURE RUNTIME· DURABLE STATE · FLEET-SCALE PLACEMENT0.97
24. OpenAl0.96
25. TRACK 1· JULY 1,20260.95
26. World's Fair0.99
27. Sandbox & Platform Engineering0.99
9:32 #23 done28 line(s)

shot 23·sharpness 2211.0
1. AlEngineer0.99
2. RUNTIME1.00
3. World'sFair1.00
4. Runtime from first principles:1.00
5. linux execution model1.00
6. Userspace1.00
7. PRESENTED BY0.96
8. Process 10.98
9. Process 20.97
10. Program instructions1.00
11. Microsoft1.00
12. Thread1.00
13. Thread1.00
14. mov eax, 11.00
15. xor ebx, ebx1.00
16. in 0x800.98
17. task_struct1.00
18. task_struct1.00
19. ioctl0.99
20. syscall0.99
21. /dev/kvm1.00
22. /dev/sda11.00
23. Linux kernel-privileged access1.00
24. Kernel gates access to hardware0.99
25. OpenAl0.97
26. TRACK 1· JULY 1, 20260.95
27. World's Fair0.99
28. Sandbox & Platform Engineering1.00

Transcript

520 cues· 7,278 words· 39,152 chars

chunk 0 · 0:12–0:53 · 114 words · 631 chars
0:12 Welcome, everyone.
0:13 Can you guys hear me OK?
0:15 I've been standing here for 15 minutes without saying anything, so we can start now.
0:20 My name is Abhishek.
0:21 I'm on the RL and agent infrastructure team at OpenAI.
0:25 What that means is we work on the infra for reinforcement learning, specifically.
0:30 And on the product side, we also develop infra that helps run untrusted code as part of ChatGPT, CodexWeb, securely and reliably at scale.
0:40 This talk is called From Fork to Fleet, Designing an Agent Sandbox Cloud.
chunk 1 · 0:46–1:28 · 120 words · 645 chars
0:46 I'll be very clear that there are a lot of words in the title that have OS and infra concepts, but this is a first principles talk.
0:53 So we'll cover what sandboxes are and why they are needed from first principles.
0:58 We will also try to cover design intuitions around designing an agent sandbox cloud to run sandboxes securely and reliably at scale.
1:07 So if some of these words don't mean anything, don't be worried.
1:10 We'll explain from first principles and go from there.
1:16 Last year, I gave a talk called How to Build an AI Sandbox from Scratch.
chunk 2 · 1:21–2:05 · 123 words · 680 chars
1:21 If you're interested, you can look at that talk as well.
1:24 Think of this as a spiritual sequel to that talk.
1:31 OK, so now let's forget about sandboxes or clouds for a second.
1:34 Let's just go back in time.
1:36 ChatGPT came out.
1:37 It's a very large pre-trained model.
1:40 People ask all sorts of questions, and it responds really, really well, really, really human-like answers.
1:45 But when people ask questions like, what is 3 plus 3, or how many R's in strawberry, sometimes it works and sometimes it doesn't.
chunk 3 · 1:54–2:38 · 125 words · 687 chars
1:54 It answers questions like what is 3 plus 3 quite well, because it's trained on the entire internet.
2:00 And apparently, people have written 3 plus 3 equal to 6 many, many times on the internet.
2:06 So it gets it right.
2:07 But people haven't asked how many hours in Strawberry enough on the internet.
2:11 And so it gets it wrong.
2:14 And so it's obvious that for anything code or math related or any problem which has a verifiable reward, which means that it can be tested whether it's true or false, the model needs something more.
chunk 4 · 2:26–3:09 · 120 words · 626 chars
2:26 And the key unlock was that given the model tool calling capability or a way to execute code, the model gets these verifiable reward questions around code and math correctly.
2:39 And that's when first principles came that if we give the models the ability to execute code, it can hill climb and be very, very good at math, code, and other domains that have verifiable rewards.
2:51 Cut to 2026, and we are seeing the consequences of doing that at scale.
chunk 5 · 2:58–3:39 · 128 words · 666 chars
2:58 So now, how can it answer what is 3 plus 3?
3:01 And how can it answer how many R's in strawberry?
3:04 Well, it can write code to do this.
3:06 So if you see the diagram, we have a training loop.
3:10 And the training loop gives it tasks or questions.
3:13 And then the training and the harness then parses the response of the model.
3:17 And the response might say, hey, execute code on my behalf.
3:22 The harness is responsible for executing the code.
3:25 And then a grader judges whether the answer is correct or not.
chunk 6 · 3:28–4:12 · 127 words · 684 chars
3:28 And then the training loop backdrops and changes the weights.
3:31 And that's how we train it to do two things.
3:34 We train it to call code execution or tools on certain classes of problems.
3:39 And secondly, we ensure that the code it executes actually solves the problem.
3:44 So this is why pool calling is important on the training side.
3:51 Now let's talk about the product side.
3:54 In the previous slide, we showed how the models are trained to emit code in order to solve certain tasks.
chunk 7 · 4:01–4:44 · 125 words · 671 chars
4:01 The agent executes the code, and we verify the reward.
4:04 Well, all of this is useless if we don't support this model on the product side.
4:08 So it's basically the same slide as before, but we don't have a training loop.

Chapters

0:00 Introduction and motivation for AI agent sandboxes
1:31 Why AI models need tools and execution environments
3:51 Product-side challenges: Security and the need for sandboxing
6:44 Comparing research vs. product sandbox requirements
8:24 Overview of the three pillars: Runtime, Persistence, and Orchestration
9:05 First principles of Linux execution: System calls and security vectors
11:15 Evaluating fork() and exec models
12:06 Understanding containers: Namespaces and cgroups
16:26 GVisor as an application kernel alternative
18:29 Hardware-level virtualization (Virtual Machines)
20:34 How VMMs (Virtual Machine Monitors) work with KVM
23:16 Evolution of modern VMMs and Rust-based safety
24:32 What defines a "microVM"?
25:43 Orchestrating microVMs via APIs
27:16 Trade-offs of microVMs (performance vs. security)
30:05 The need for persistent storage in agent sandboxes
31:40 Use cases for persistence: Reliability, long-running tasks, and research
34:36 Design choices for disk snapshotting
36:03 First principles of Linux block storage and file systems
37:25 Implementing always-on vs. explicit persistence
41:20 Scaling and orchestrating sandboxes at fleet level