When people say "just point Opus/Codex at localhost using Playwright MCP," they've obviously never attempted to do exploratory testing by leveraging parallel browser workers against stateful applications. A coding agent running in a loop spends almost all of its context window struggling with DOM selectors and doesn't even detect state persistence problems like the 2.5h->25h decimal inflation upon refreshing the page as demonstrated in the demo.
Managing read/write resource leases for parallel sub-agents not to step on each other's test fixtures is already a non-trivial orchestration challenge by itself.
replyAgreed! We've also found that the tooling you use makes a big difference in how efficient and precise the agent is, which is why we included the comparison of Playwright MCP vs agent-browser. When we were first starting Mo we tried using several different browser interfaces but we weren't able to get the recall we wanted. We settled on rewriting some custom tooling based on our mcp for our investigatory/exploratory agents. Tooling about how to interact with the A11y and DOM, breaking out easy CUA subproblems, and browser infra are still all non-trivial problems.
replyjust have engineers not sleep at the wheel and implement appropiate solutions when agent struggles?
reply