Marco Montanari's MIT MCP server (one stdlib Python file) creates a git folder, starts claude --remote-control in screen, and lets you follow from claude.ai/code, optional code-server, or Telegram. It ships a storage-free OAuth 2.1 layer for claude.ai connectors, bracketed paste into the TUI, and Stop hooks that relay replies and permission asks.
Subquadratic's public preview Snap scores bounded answers (yes/no, choice, score) over up to 384K tokens of text and images in about a second at 100K, so coding agents can offload file selection, tool choice, and test verification while leaving generation to a frontier LLM. It averages 67.6 vs Jev's 61.6 on 16 enterprise Decision Index tasks, costs $0.05 per million input tokens with free outputs, and uses Subquadratic Sparse Attention instead of quadratic prefill…
Signals (public preview) watches Managed Agents and Inference Engine traffic for interaction, tool execution, and infra failures, then one click exports flagged sessions into Evaluations scenarios. Simulations replays those multi turn conversations in Harness Runtime sandboxes against Claude Code, Codex, or OpenCode candidates so you can grade a fix before users see it…
Coleman Parkes polled 300 senior leaders at large US and UK mission critical shops for Undo and found 81% had an AI tool linked production incident or outage in the prior six months, with engineers averaging 16.9 hours a week debugging versus 9.8 writing code. 35% of AI generated code reached production before teams fully understood it, and 79% said release cycles were no faster…
Muse Wonder is live today on the Meta Model API and Muse Code (Wonder Full for Muse Code subscribers), aimed at personal and professional agents with Active Memory, Active Focus, and frontier coding for building apps and games. Meta says both models lead on agent, memory, and coding benchmarks and will soon power more of Muse for subscribers first…
Across 6,840 synthetic call graph refactoring tasks, six agents in Copilot's harness stayed near perfect by scripting around size with tools like grep , while plain LLMs degraded; on SWE bench Verified traces, the count of read and analyze tool calls predicted failures better than lines edited. A new task that merges two near identical classes finally broke the agents, and code plus data are open at github.com/microsoft/CABRA…
Spotted something we missed? Start a thread.
