Browserbase: Fully Autonomous Is The Wrong Goal, 99% Accurate AI Agents Still can fail catastrophically
Shoal Signal with
@derekmeegan, Engineering Lead at @browserbasehq. Hosted by
@zaddycoin.
We get into why a browser agent that is 99% accurate per step still fails most of the time, terminal value vs continuous risk, success per unit of work, why security is the root of scalability, and why fully autonomous is the wrong goal.
0:00 Intro and what a browser agent is
0:33 Browser agents vs. Playwright scripts
6:33 Three types of browser trajectories
8:48 Where automation breaks down, and cost
11:12 Real use cases and the Ramp receipt flow
14:23 Terminal value, continuous risk and cost
18:11 Next token prediction and removing steps
23:55 Success per trajectory vs. per unit of work
29:36 Harnesses, the human, and protecting context
36:11 Humans and agents as complements, not a binary
42:07 Sci-fi vs. reality, it is matrix multiplication
48:31 Compliance, bill pay, and infra at scale
50:45 Agent security, OpenClaw, and credential brokering
1:00:46 Browser-to-API and the power of skills
1:11:15 The real alpha: upskill yourself