Reality Check: Agent Benchmarks, Browser Wars, and the Reliability Tax

Analysis

New Research Quantifies Where Agents Actually Fail

A growing body of work is finally moving past 'agents are cool' benchmarks toward measuring specific failure modes like tool misuse and context drift across long horizons. This matters because founders shipping agents need failure taxonomies, not just leaderboard scores, to build real guardrails.

Tools

Browser-Use Gains Traction as the Open Alternative to Proprietary Browser Agents

With big labs locking down their own browser agent stacks, open source projects like this are becoming the default for builders who don't want vendor lock-in on something this core. Expect forks and enterprise wrappers to multiply fast now that it has real momentum.

News

MCP Spec Update Tackles Auth and Session Persistence

The Model Context Protocol keeps patching the boring but critical stuff, auth flows and session state, that determines whether agents can actually operate across tools reliably in production. This is the plumbing work nobody tweets about but everyone building real agent systems is waiting on.