Coding agents & tooling releases

• Sonnet 5 live on B.AI's API — dual-routed (official + custom), 1M context, adaptive thinking, big agentic-coding gains.
• Claude Code v2.1.198 ships Claude-in-Chrome GA, background-agent Notification hooks (agent_needs_input/agent_completed), and a /dataviz skill.
• Fable 5 is back with rate limits reset; theo's tips for avoiding limits — use "high" effort (xhigh/max burn tokens for worse output), delegate implementation to Codex/GPT-5.5 as fallback.
• Kimi K2.7 Code GA in GitHub Copilot — first open-weight model in the Copilot picker, hosted on Azure.
• Copilot vision GA — attach images/PDFs so Copilot reasons over what it sees alongside code.
• Enterprises can default Copilot to auto model selection via managed-settings.json.

Multi-agent & orchestration patterns

• Bloome multi-agent group chat — Claude Code, Codex, DeepSeek, and a lead agent debate/build in one shared thread, no copy-paste context loss.
• MCP server for two Claude Code agents to pair-program directly, no human relaying messages.
• Reducing usage via lower-effort subagents when the main session runs High/XHigh — cost-aware routing in practice.
• Proposed /research skill — background agent scans high-trust sources, saves findings to markdown.
• Skill for planning large work — maps decision "fog of war," scopes each decision to fit its own parallel context window.

Benchmarks & eval

• Senior SWE-Bench evaluates agents on realistic, under-specified tasks instead of junior-style over-specified ones; HN pushback: subjective LLM grading undermines rigor.
• DeepSpec — full-stack training/eval codebase for speculative-decoding draft models.

Agentic SaaS & enterprise deployment

• Cursor's Forward Deployed Engineers set up agent "software factories" inside enterprises.
• Warp CEO on software factories as the next phase — CLI tool evolving into an automated build platform.
• Autoresearch / self-improving agent loops — "outer loop" agents maintain and improve other agents, humans stay central.
• Browser tools for GitHub Copilot in VS Code GA with scoped permissions and network domain controls.
• Tabstack — hosted web automation for apps/agents, no browser to run yourself.
• CubeSandbox — RustVMM/KVM-based hardware-isolated sandbox for agents, E2B SDK-compatible.

Data & retrieval tooling

• olmocr — converts PDFs/images to clean Markdown for LLM training data, handles tables/equations/handwriting.
• Folder-to-queryable-with-citations tool compiles a data folder so an agent can query it with sourced citations.