--- schema: 1 type: video title: 'Bringing agents onto the world wide web — Paul Klein IV, Browserbase' description: >- When OpenClaw shipped, people started buying Mac minis to run it from home, SSHing in and clearing captchas off a residential IP. Paul Klein IV points out he has yet to see a SOC 2 compliant Mac Mini setup at scale, and that this felt like a reasonable answer is itself the problem. His argument is that browser agents are no longer held back by the models. The capability is already there and the engineering around it is missing, which makes the overhang something any team can close rather than wait on a lab for. That engineering has three parts. The most reliable browser agents in production are multimodal and write code alongside clicking, often intercepting network requests and replaying them rather than driving pixels. They carry a real harness, with skills and memory so a site is not rediscovered every run, and with page context compressed rather than dumped whole into the model. And they sit on infrastructure that renders a page identically every time, since a layout that comes back mobile on one run and desktop on the next produces results the agent cannot account for. He then turns to what the web owes agents: accessibility trees, Chrome's new Web MCP, and two unsolved problems, how an agent logs in on your behalf and who certifies that an agent can be trusted. The payoff is not in San Francisco. It is the logistics company in Singapore, the bank in South Africa, and the lumber factory in Mexico, all running on PHP forms with people clicking buttons every day. Speaker info: - https://x.com/pk_iv - https://www.linkedin.com/in/paulkleiniv/ - https://github.com/browserbase/stagehand Timestamps: 0:00 - Why web agents have not happened yet, and is it the models 3:17 - The missing piece is the harness 4:32 - Harnesses that beat the baseline model 6:01 - The capabilities overhang in computer use 7:05 - Multimodal agents, skills, and token efficiency 9:09 - Infrastructure, and the SOC 2 Mac Mini problem 10:24 - What the web owes agents: accessibility and Web MCP 11:40 - Authentication, trust, and who issues the certificate 14:07 - What a real platform has to provide 15:45 - The real economy runs on PHP forms canonical: 'https://liquidgold.cc/watch/GqoNrUz8hEU' markdown: 'https://liquidgold.cc/watch/GqoNrUz8hEU.md' llms: 'https://liquidgold.cc/watch/GqoNrUz8hEU/llms.txt' parent: 'https://liquidgold.cc/channel/aidotengineer.md' channel: aidotengineer updated: '2026-10-24T05:24:33.001Z' language: null children: [] videoId: GqoNrUz8hEU durationSeconds: 1106 publishedAt: '2026-08-14T15:00:31.000Z' searchChunks: - ordinal: 0 text: null startSeconds: 24 endSeconds: 199 watch: 'https://liquidgold.cc/watch/GqoNrUz8hEU?t=24' heldReason: rights - ordinal: 1 text: null startSeconds: 24 endSeconds: 199 watch: 'https://liquidgold.cc/watch/GqoNrUz8hEU?t=24' heldReason: rights - ordinal: 2 text: null startSeconds: 24 endSeconds: 199 watch: 'https://liquidgold.cc/watch/GqoNrUz8hEU?t=24' heldReason: rights - ordinal: 3 text: null startSeconds: 209 endSeconds: 279 watch: 'https://liquidgold.cc/watch/GqoNrUz8hEU?t=209' heldReason: rights - ordinal: 4 text: null startSeconds: 279 endSeconds: 290 watch: 'https://liquidgold.cc/watch/GqoNrUz8hEU?t=279' heldReason: rights - ordinal: 5 text: null startSeconds: 367 endSeconds: 379 watch: 'https://liquidgold.cc/watch/GqoNrUz8hEU?t=367' heldReason: rights - ordinal: 6 text: null startSeconds: 435 endSeconds: 447 watch: 'https://liquidgold.cc/watch/GqoNrUz8hEU?t=435' heldReason: rights - ordinal: 7 text: null startSeconds: 447 endSeconds: 554 watch: 'https://liquidgold.cc/watch/GqoNrUz8hEU?t=447' heldReason: rights - ordinal: 8 text: null startSeconds: 447 endSeconds: 554 watch: 'https://liquidgold.cc/watch/GqoNrUz8hEU?t=447' heldReason: rights - ordinal: 9 text: null startSeconds: 447 endSeconds: 554 watch: 'https://liquidgold.cc/watch/GqoNrUz8hEU?t=447' heldReason: rights - ordinal: 10 text: null startSeconds: 565 endSeconds: 632 watch: 'https://liquidgold.cc/watch/GqoNrUz8hEU?t=565' heldReason: rights - ordinal: 11 text: null startSeconds: 642 endSeconds: 709 watch: 'https://liquidgold.cc/watch/GqoNrUz8hEU?t=642' heldReason: rights - ordinal: 12 text: null startSeconds: 642 endSeconds: 709 watch: 'https://liquidgold.cc/watch/GqoNrUz8hEU?t=642' heldReason: rights - ordinal: 13 text: null startSeconds: 642 endSeconds: 709 watch: 'https://liquidgold.cc/watch/GqoNrUz8hEU?t=642' heldReason: rights - ordinal: 14 text: null startSeconds: 709 endSeconds: 720 watch: 'https://liquidgold.cc/watch/GqoNrUz8hEU?t=709' heldReason: rights - ordinal: 15 text: null startSeconds: 720 endSeconds: 853 watch: 'https://liquidgold.cc/watch/GqoNrUz8hEU?t=720' heldReason: rights - ordinal: 16 text: null startSeconds: 720 endSeconds: 853 watch: 'https://liquidgold.cc/watch/GqoNrUz8hEU?t=720' heldReason: rights - ordinal: 17 text: null startSeconds: 864 endSeconds: 950 watch: 'https://liquidgold.cc/watch/GqoNrUz8hEU?t=864' heldReason: rights - ordinal: 18 text: null startSeconds: 960 endSeconds: 1106 watch: 'https://liquidgold.cc/watch/GqoNrUz8hEU?t=960' heldReason: rights --- # Bringing agents onto the world wide web — Paul Klein IV, Browserbase When OpenClaw shipped, people started buying Mac minis to run it from home, SSHing in and clearing captchas off a residential IP. Paul Klein IV points out he has yet to see a SOC 2 compliant Mac Mini setup at scale, and that this felt like a reasonable answer is itself the problem. His argument is that browser agents are no longer held back by the models. The capability is already there and the engineering around it is missing, which makes the overhang something any team can close rather than wait on a lab for. That engineering has three parts. The most reliable browser agents in production are multimodal and write code alongside clicking, often intercepting network requests and replaying them rather than driving pixels. They carry a real harness, with skills and memory so a site is not rediscovered every run, and with page context compressed rather than dumped whole into the model. And they sit on infrastructure that renders a page identically every time, since a layout that comes back mobile on one run and desktop on the next produces results the agent cannot account for. He then turns to what the web owes agents: accessibility trees, Chrome's new Web MCP, and two unsolved problems, how an agent logs in on your behalf and who certifies that an agent can be trusted. The payoff is not in San Francisco. It is the logistics company in Singapore, the bank in South Africa, and the lumber factory in Mexico, all running on PHP forms with people clicking buttons every day. Speaker info: - https://x.com/pk_iv - https://www.linkedin.com/in/paulkleiniv/ - https://github.com/browserbase/stagehand Timestamps: 0:00 - Why web agents have not happened yet, and is it the models 3:17 - The missing piece is the harness 4:32 - Harnesses that beat the baseline model 6:01 - The capabilities overhang in computer use 7:05 - Multimodal agents, skills, and token efficiency 9:09 - Infrastructure, and the SOC 2 Mac Mini problem 10:24 - What the web owes agents: accessibility and Web MCP 11:40 - Authentication, trust, and who issues the certificate 14:07 - What a real platform has to provide 15:45 - The real economy runs on PHP forms Canonical: https://liquidgold.cc/watch/GqoNrUz8hEU · Channel: [AI Engineer](https://liquidgold.cc/channel/aidotengineer.md) > Search chunk text is held until public-source provenance is verified. ## Document chunks > Missing: no stored document chunks are available.