Claude (Opus / Sonnet / Haiku), Claude Code as the daily agentic interface, MCP as the tool-integration standard, Computer Use as the desktop-control pattern. The lab that introduced Constitutional AI and shipped the most opinionated take on AI-native software engineering.
Opus, Sonnet, Haiku — three points on the capability/cost curve
Claude Opus for orchestration and complex architecture decisions, Sonnet for bulk production workloads with strong reasoning, Haiku for fast cheap classification and routing. Native tool use, computer use, agent SDK, prompt caching, extended thinking. Constitutional AI training produces noticeably different default behavior — less sycophantic, more willing to push back on flawed premises.
Anthropic's agentic CLI for software engineering
The terminal-resident agent that has replaced ad-hoc Claude API calls for most of my development work. Hooks, slash commands, sub-agents, MCP server integration, plan mode, the Tasks system. Where Cursor wins on IDE integration, Claude Code wins on agentic depth — multi-step workflows that span hours, with proper context management and tool use.
The N×M → N+M tool-integration standard
Anthropic-introduced open protocol for connecting LLM apps to tools and data sources. The architectural reduction: instead of every agent re-implementing every tool, every tool exposes itself once via an MCP server and every agent connects via MCP client. The same pattern that USB-C did for hardware peripherals, applied to AI tool integration.
Claude operating a desktop via screenshots + actions
Vision + tool calling fused into a desktop-control loop. Claude looks at screenshots, decides on clicks/keypresses/scrolls, executes via a coordinated agent harness. Beta but production-shipped in specific verticals. The architectural pattern matters more than the current accuracy: it generalizes to any "drive a UI" task where API access is unavailable.