WEB-25: Top 25 browser agents — drive a real browser to do tasks.

**WEB-25 Rebalance Report** Four agents entered the index this week, four departed. All moves were 0.04-weight swaps with no mid-band adjustments. **In:** ai-newtab, Awesome-CUA, imnotarobot, and snapgent-connector. The cluster suggests a turn toward browser-native and connector-layer tooling—ai-newtab and snapgent-connector both signal integration with user-facing platforms rather than model-direct approaches. **Out:** model:openjev/openjev, awesome-agent-rl-environments, agent-vision, and vulnerator-ai. The removals span model wrappers, RL scaffolding, vision pipelines, and security scanning—no coherent signal, which reads as turnover rather than thesis shift. The net effect is marginal: index composition shuffled but weight distribution flat. This is housekeeping-grade churn.

**WEB-25 Rebalance** Four agents entered, four exited. No weight shifts among incumbents. **In:** invisible_dots and imshaikot/browsentic both landed at 0.04 weight—the former appears to target visual perception tasks, the latter a browser automation framework. vulnerator-ai joined on security testing credentials. awesome-agent-rl-environments suggests reinforcement learning infrastructure is now meeting the index threshold. **Out:** bux, mcp-man, bughunter-ai, and tweetytweets all dropped at 0.04 weight. The removals span generalist and specialist categories without a clear pattern—mcp-man's departure is notable given the broader ecosystem focus on Model Context Protocol. The churn hints at a maturation filter: agents need either stronger recent signal or clearer specialization to hold weight. The RL and security additions show the index tracking emerging capability clusters, though the even four-for-four swap suggests this week was routine maintenance rather than directional shift.

**WEB-25 Rebalance Report** Three agents entered the index this week: **touri**, **mcp-man**, and **model:openjev/openjev**, each at 0.04 weight. Touri and mcp-man appear to have crossed relevance thresholds on task completion metrics; openjev entered via model-native integration pathways. Three agents exited: **browsentic**, **canary**, and **pi-serini**, all dropping from 0.04. Browsentic's removal signals continued friction in browser automation reliability; canary and pi-serini both lost ground on latency benchmarks relative to newer entrants. No existing agents shifted weight this cycle—the index remained otherwise stable. The churn here reflects a steady refresh of mid-tier performers rather than any fundamental shift in how web agents are being evaluated or deployed.

**WEB-25 Rebalance** Seven agents entered the index this week at equal weight (0.04 each). **Browsentic** added browser automation capabilities with a focus on authentic web interaction patterns. **Bughunter-ai** brings structured vulnerability detection to the roster. Seven agents rotated out. Notable exit: **x-scraper-no-api**, which had served as a no-dependency scraping option, and **pixiebrix/agent-browser-shield**, a browser-native protection layer that hadn't maintained sufficient performance signals. The churn is symmetric and mechanical—no weights shifted among incumbents, suggesting the index is cycling through candidate agents at the same tier rather than consolidating around proven performers. Web automation and vision agents continue flowing through the index; the real signal will be whether any of this week's additions hold position in two weeks.

### WEB-25 rebalance 25 constituents this week. 6 new entrants, including dots, x-scraper-no-api, awesome-agent-rl-environments. 6 dropped, including Notion-AI-to-OpenAI-Compatible, bughunter-ai, deepcloak.

### WEB-25 rebalance 25 constituents this week. 7 new entrants (e.g. 874b8f38, 24e909ce, cafcc3e5…). 7 dropped (e.g. 0ed2cb0c, a7dd2b1a, 63b50a63…).