A teardown of mobile agent boxes finds Claude Code on Anthropic’s own Firecracker microVM (custom Rust PID 1, persistent disk) while Instinct rents E2B Firecracker sandboxes and keeps durable memory as a git Markdown vault in S3. Same isolation primitive; different durability and where the harness lives.
With Theorem, Claude patched the bugs it found and then produced about 180 million lines of Rocq proving those fuzz targets never crash on any input under 64 MB, checked by the Rocq kernel rather than human reviewers. The guarantee only covers single threaded static RISC V Linux binaries, and a subset of proofs plus a Claude skill for verifying your own code are public (proofs)…
Speaking at Grafana's ObservabilityCON, Katelyn Lesse said her roughly 200 person team, which runs Anthropic's core services and developer tools, no longer writes code by hand: engineers own system design and verification while Claude Code does the implementation, including a Python to Rust migration. Some test fixes already run with no human in the loop, and Anthropic is weighing letting agents change real codebases on their own…
Launched Oct 23, v2.0 averages Cognition's private FrontierCode v1.1 (graded on whether a maintainer would merge the change), Terminal Bench 4.0, and re graded SWE Atlas QnA AA across 290 tasks, so scores reset below v1.5. Claude Code with Opus 5.5 tops it at 61.7 for $11.42 per task, while Codex with DeepSeek V4.1 Flash scores 47.3 for $0.39 (announcement)…
Announced Oct 21 for late October, the Code tab in the Copilot app on Windows, Mac, and web builds trackers, dashboards, apps, and scripts from prompts grounded in your work data, then publishes them to Copilot Managed Runtime to share inside your org. It needs Frontier enrollment, a Microsoft 365 Copilot license, and a usage based spending policy, and local coding with your files and terminal stays off until an admin enables it (message center)…
Released Oct 23, it saves reusable prompts and slash commands in ten numbered /presets slots, lets you set model, reasoning effort, and permission profiles inline with /model and /permissions , and authenticates enterprise MCP connections via Codex account grants or dedicated sign in. It also masks credentials in remote Bash and Zsh snapshot sessions and stops failed compaction from leaving partial summaries in history…
Spotted something we missed? Start a thread.
