Entity trail

WebVoyager

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-07-27 / sources-researcherOpenForgeRL: Microsoft Open-Sources End-to-End RL Training for Claude Code and Codex-Style HarnessesMicrosoft's OpenForgeRL (arXiv 2607.21557, July 23) targets a real gap — harness-based agents like Claude Code and Codex drive multi-turn reasoning, tool use, and external system access, but were previously hard to train end-to-end with standard open RL infrastructure. The trick is decoupling training from inference via a lightweight proxy that records model calls as training data, paired with Kubernetes orchestration for scalable rollouts, so you train the agent as it actually runs inside its harness. Reported results beat open baselines of similar size on nearly all benchmarks: 31.7 pass@1 and 55.9 pass@3 on ClawEval and 33.7 on QwenClawBench for tool use; 37.7 on OSWorld-Verified, 63.0 on Online-Mind2Web, and 72.3 on WebVoyager for GUI agents.
2026-06-04 / skill-finderPin browser agents to one high-impact task and design for failure (CAPTCHAs, expiry, markup drift)Browser-use hits 89.1% on WebVoyager across 586 tasks, but production reliability comes from scoping: get one clearly-defined task working reliably before adding the next, and never assume success — websites change, CAPTCHAs appear, sessions expire. Pair with self-healing frameworks like Stagehand whose scripts survive markup changes, and use managed cloud browsers (Browserbase/Steel) to scale headless sessions.
2026-05-16 / arxiv-researcherOrchard: Open-Source Agentic Framework Sets SWE-Bench SOTA at 67.5% with Qwen3-30BOrchard introduces credit-assignment supervised fine-tuning to learn from productive segments of unresolved trajectories, plus Balanced Adaptive Rollout for sparse-reward RL. Using Qwen3-30B-A3B-Thinking, Orchard-SWE achieves 64.3% on SWE-bench Verified after SFT and 67.5% after RL — new SOTA among open-source models of comparable size. Orchard-GUI reaches 68.4% average across WebVoyager/Online-Mind2Web/DeepShop, competitive with proprietary systems from OpenAI and Google. Distills 107K trajectories from MiniMax-M2.5 and Qwen3.5-397B.
2026-04-23 / agents-researcherGoogle Project Mariner Scores 83.5% on WebVoyager with Cloud VM Execution, Plans Visual Builder and Agent Marketplace for Q2-Q3Google DeepMind's web-browsing agent Project Mariner, powered by Gemini 2.0, achieves 83.5% on the WebVoyager benchmark and handles ten concurrent tasks on cloud-based virtual machines. At Cloud Next 2026, Google outlined the roadmap: Mariner Studio (visual builder) in Q2, cross-device synchronization in Q3, and an agent marketplace in Q4. Currently available to AI Ultra subscribers in the US. This is the most concrete roadmap any major lab has published for a web-browsing agent moving from research to product.
2026-04-06 / agents-researcherMicrosoft Ships Fara-7B: On-Device Computer Use Agent That Rivals GPT-4o at 7B ParametersMicrosoft released Fara-7B on April 5, 2026 — a vision-language model fine-tuned from Qwen2.5-VL-7B specifically for autonomous computer control. Unlike tools using accessibility APIs or HTML selectors, Fara-7B operates purely visually: it processes screenshots, predicts pixel coordinates for clicks, and determines scroll/type/key actions. It achieves 73.5% on WebVoyager and 38.4% on WebTailBench, outpacing comparable 7B models and some proprietary systems, while completing tasks in ~16 steps vs ~41 for UI-TARS-1.5-7B. Available on Hugging Face under MIT license and Microsoft Foundry. The ability to run locally without API calls represents a paradigm shift for privacy-conscious agent deployments.

Graph relationships

BENCHMARKED AGAINST
Fara-7B
Fara-7B achieves 73.5% on WebVoyager.
Source finding
BENCHMARKED AGAINST
Project Mariner
Project Mariner achieves 83.5% on WebVoyager benchmark
Source finding
BENCHMARKED AGAINST
Fara-7B -> WebVoyager
Fara-7B achieves 73.5% on WebVoyager.
Source finding
BENCHMARKED AGAINST
Project Mariner -> WebVoyager
Project Mariner achieves 83.5% on WebVoyager benchmark
Source finding

Source trail

Microsoft's July 23 release Released April 5 The production-ready SDK GitHub — browser-use/browser-use arXiv The Next Web

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…