AgentWebBench: Benchmarking Multi-Agent Coordination in Agentic Web

AgentWebBench addresses a specific evolution of the web: the transition from human-centric browsing to agent-to-agent (A2A) interaction. While existing benchmarks like WebArena and Mind2Web focus on a single agent navigating a DOM, this project introduces the 'Content Agent'—a site-specific proxy that mediates data access. Defensibility is low (3) because, as a 4-day-old research project with 0 stars, it lacks the 'gravity' or adoption required for a benchmark to become a standard. Its value resides entirely in the methodology and the specific evaluation datasets. Frontier risk is medium because while OpenAI and Anthropic are obsessed with 'Computer Use' (UI-based), they are also defining the protocols for data exchange (e.g., Anthropic's Model Context Protocol). If a frontier lab releases a standardized A2A protocol, this benchmark might be rendered obsolete unless it pivots to evaluate that specific protocol. Platform domination risk is high; the 'Agentic Web' is currently a battleground where Microsoft, Google, and OpenAI are building the infrastructure. These players are likely to release their own evaluation suites to steer the industry toward their preferred coordination patterns. The 3 forks relative to 0 stars indicate early interest from the academic community, but it faces stiff competition from established benchmarks like GAIA or BigBench.