AI & Automation Briefing - June 30, 2026
OSWorld 2.0 Shows Frontier AI Agents Complete Just 20% of Complex Real-World Tasks
Researchers released OSWorld 2.0, a benchmark of 108 long-horizon computer-use workflows designed to reflect the complexity of real professional tasks. Each workflow takes human users a median of 1.6 hours and requires an average of 318 tool calls. Claude Opus 4.8 with maximum thinking scored best under the benchmark's primary metric, completing only 20.6% of tasks with a 54.8% partial score. GPT-5.5 was more token-efficient but topped out near 13%. The benchmark found that agents don't typically fail on basic GUI control. They fail by losing track of constraints mid-task, missing information that arrives during execution, guessing instead of querying the user, and skipping verification steps. These failure patterns map directly onto the weak points of automated pipelines built on tools like n8n and Claude, and make a concrete case for keeping human checkpoints in any long-horizon agentic workflow.
35B Agent Matches Trillion-Parameter Models by Scaling Task Depth, Not Size
Shanghai AI Lab released Agents-A1, a 35B mixture-of-experts model that competes with trillion-parameter systems by training on long-horizon knowledge-action trajectories rather than simply increasing parameter count. The model learns by chaining reasoning, tool use, execution feedback, and verification steps together, with knowledge distilled from multiple specialized teacher models. It performs competitively on benchmarks covering scientific reasoning, research-level coding, ML engineering, and scientific tool use, including HLE with tools, SciCode, and MLE-Bench-Lite. For teams building multi-agent workflows, this is a concrete signal that routing tasks to a well-trained, task-depth-optimized smaller model can outperform relying on raw model scale. Agents-A1 is released as an open model.
Cursor Launches Mobile App for Remote Coding Agent Oversight
Cursor released a mobile app that lets users start new coding agents or continue conversations with agents already running on their desktop client. The app builds on Cursor 2.0, released in October, which repositioned the tool around autonomous coding agents rather than direct code editing. The move follows similar mobile releases from Anthropic and OpenAI, and reflects a wider pattern in AI coding tools: developers are shifting from complex desktop setups toward async, phone-based oversight of agents doing the actual coding work. Anthropic's head of Claude Code, Boris Cherny, noted at a recent talk that he now does most of his coding on his phone, something he said he would have dismissed as unrealistic six months ago. For automation builders, this signals that managing long-running agents remotely is becoming a core operational skill, not a workaround.