Summary
Paul Klein IV, founder of Browserbase, argues at AI Engineer that web agents are not being held back by model capabilities — models have improved dramatically in the past year, with significant RL investment in computer use environments — but by missing infrastructure and harness engineering. Drawing on Andrej Karpathy’s 2023 framing of the ideal LLM system (code interpreter, browser, audio/video input, sub-agents), Klein traces how those predictions have largely come true, yet production web agents still routinely fail on real websites.
The core thesis is that an “agent harness” — the scaffolding surrounding the model — is now where most of the leverage lies. Klein breaks this down into practical components: giving agents memory and reusable skills so they don’t rediscover site behavior on every run, presenting optimized token representations of web pages rather than dumping raw HTML, and combining browser control with code-writing agents for tasks where scripted replay is more reliable than live inference. He cites the example of Claude Code outputting Playwright scripts rather than directly driving Chrome as evidence that hybrid approaches outperform pure browser use.
Klein also addresses the infrastructure layer: consistent, SOC 2-compliant browser environments that can scale to thousands of parallel agents are unsolved, as illustrated by the community workaround of running OpenCloud on home Mac Minis for trusted IP addresses. Browserbase’s browser.sh product, which publishes pre-built skills for common websites via WebMCP, is presented as one approach to reducing per-task discovery overhead and improving token efficiency at scale.
📺 Source: AI Engineer · Published August 14, 2026
🏷️ Format: Deep Dive
