Datadog made Code Execution generally available on its MCP Server: agents run sandboxed JavaScript to query logs, metrics, and traces in parallel, then return only joined evidence. Across 25 tasks it claims +15.6 pp correctness, 73.2% fewer input tokens, and 39.6% fewer tool calls versus the core toolset.
Peter James asked Muse to zip its visible files to Google Drive and got 2.7GB compressed ( 6.8GB unpacked) of the Hatch Linux session root—SOUL/IDENTITY markdown, 68 skills, Spaces templates, unused Codex CLI 0.149.0, and SSH key files. Meta’s bug bounty marked it Not Applicable; no sandbox escape or cross user access was shown…
GitHub Copilot for JetBrains 1.18.0 adds public preview assisted approvals (auto ok low risk tool calls), Codex plan mode before implementation, org/enterprise shared skills, and persistent per tool MCP controls including a toggle for the built in GitHub MCP server…
Artificial Analysis ranks Claude Opus 5.5 (max) at 58 on its Intelligence Index—the highest it has measured—tying GPT 6 Astra on Terminal Bench 4.0 (59.6%) and leading six of ten evals. Anthropic also cut Opus to $4/$20 per M tokens and cache reads to $0.20…
Claude Code 2.1.280 switches Pro ($20/mo) and Team Standard defaults from Sonnet to Opus, matching Max and Enterprise, and makes Opus 5.5 the default Opus model. Subscription usage still counts against the five hour/weekly allowance; switch back with /model or /config if you want Sonnet…
NVIDIA, NTU, and MIT released SoL Pi (MIT), an opt in extension for the open source Pi coding agent that cuts EdgeBench token traffic 44.7–49% and API cost about one third versus base Pi, with scores within 6%. Four harness layer mechanisms—Action Fusion, Online Context Compact, ObservationPack, and Evidence Preserving Reducer—ship on GitHub; paper at arXiv:2609.20519…
Spotted something we missed? Start a thread.