Anthropic shipped 10 ready-to-run agent templates for financial services work — pitchbooks, KYC screening, month-end close — plus Microsoft 365 add-ins for Excel, PowerPoint, and Word. Claude Opus 4.7 leads the Vals AI Finance Agent benchmark at 64.37%, and this is the first vertical where Anthropic is shipping domain-packaged agentic workflows out of the box.
On May 6, Anthropic’s first developer conference delivered six interconnected launches: a 300MW SpaceX compute deal, doubled Claude Code rate limits, Code Review GA at $15–25/PR, three Managed Agents upgrades, and an 80x Q1 growth figure that outpaced the company’s own forecast by 8×. Taken together, they describe a company that is no longer just building a model — it is building the infrastructure layer for autonomous software development.
Claude Code Review went GA on May 6: a multi-agent system that dispatches parallel bug-finders, runs a verification pass to cut false positives, and lands a single high-signal comment on your pull request. It costs $15–25 per review, is billed separately from your plan, and is available on Team and Enterprise only. Here is how the architecture works, what the performance numbers say, and how it stacks up against /ultrareview, Cursor Bugbot, Greptile, and CodeRabbit.
Novee Security disclosed a CVSS 9.9 remote code execution vulnerability in Cursor on April 28, patched in version 2.5. The attack vector — a malicious git hook triggered automatically by Cursor’s own agent — is not a bug that better code can fully solve. It is a consequence of putting an autonomous AI agent inside a process that has broad, native system access.
Anthropic’s Dreaming feature — launched at Code with Claude SF on May 6 — lets managed agents review their own past sessions overnight, curate what they learned, and arrive at the next run measurably better. Harvey, the legal AI company, saw task completion rates increase 6× after deploying it.
CLAUDE.md is the single highest-leverage file in any Claude Code project — and the most-misunderstood. This is the comprehensive guide: what to put in it, what NOT to put in it, the structure that compounds over time, and the security model that prevents CVE-2026-21852-style supply-chain attacks.
Skills, plugins, and MCP servers are the three ways to extend Claude Code — and they look similar enough that engineers routinely pick the wrong one. This is the reference: what each one is, when to reach for it, what they cost, and the failure modes nobody warns you about.
OpenAI’s March 2026 acquisition of Astral — makers of uv, Ruff, and ty — hands one AI lab control over Python’s most critical developer infrastructure. The tools stay open source, for now. The governance question is wide open.
Version 2.1.126 ships four practical upgrades: a /model picker that reads your gateway’s model list, a project purge command for clean state management, OAuth paste mode for headless environments, and an Auto mode that tells you when it’s stuck.
Direct answers to the questions developers actually ask about Claude Code in 2026 — plans, models, installation, CLAUDE.md, auto mode, MCP, parallel agents, cost, enterprise, security, and how it stacks up against Cursor, Copilot CLI, and Gemini CLI.