Two days ago Anthropic held their first developer conference, Code with Claude, and launched Claude 4. I’ve been on the Max plan since April, using Claude Code on personal projects nights and weekends. What happened Thursday changed how I think about building software.
What Actually Happened Link to heading
Claude Code is a command-line tool. Open a terminal, talk to it. It reads your files, writes code, runs tests, and understands your entire codebase. No IDE, no browser. Just your terminal. It launched in February alongside Claude 3.7 Sonnet and became generally available Thursday.
I pointed it at a personal project — hundreds of files, custom build pipelines, years of accumulated complexity. It oriented itself in seconds. Found the right files, understood the patterns, made changes that fit the existing architecture. I’ve onboarded senior engineers who took weeks to reach that level of contextual awareness.
The Claude 4 benchmarks tell the story. Opus 4 scores 72.5% on SWE-bench Verified, Sonnet 4 hits 72.7%. SWE-bench tests whether a model can take a real GitHub issue and produce a working patch. Nearly three-quarters of the time, it does.
The Productivity Shift Is Real Link to heading
After twenty years building financial systems at scale, I know the bottleneck was never typing speed. It’s context switching, understanding unfamiliar code, and keeping dozens of files in your head simultaneously. Imagine what enterprises could unlock by giving developers this kind of leverage.
Claude Code collapses that. I describe what I want, and it:
- Searches the codebase for relevant files
- Reads them to understand patterns and conventions
- Writes the implementation, matching existing style
- Runs the tests and fixes what breaks
- Commits with a coherent message
On my personal projects, an evening of focused work gets done in minutes. Not a rough draft. Production-quality code that passes CI.
CLAUDE.md deserves more attention. Drop a markdown file in your project root and the agent reads it every session — build commands, coding conventions, architectural decisions, editorial voice. The agent doesn’t just read your code. It reads your rules. Every session compounds on the last.
Where It Gets Interesting Link to heading
Extended thinking shipped with Claude 3.7 Sonnet in February. Claude 4 takes it further by interleaving thinking with tool use — the model reasons through a problem, reaches for a tool, incorporates what it finds, and keeps reasoning. This isn’t autocomplete. It’s a thinking process that produces code.
I watched it debug a race condition in a personal project’s async pipeline. Hypothesized, added logging, ran the code, read the output, revised its hypothesis, fixed the root cause. Three minutes. My debugging methodology, executed faster than I could type the commands.
Opus 4 and Sonnet 4 also execute tool calls in parallel — reading five files simultaneously instead of sequentially. Small detail, huge impact on fluidity.
GitHub Actions and Multiplying Yourself Link to heading
Claude Code’s GitHub Actions integration changed how I run my personal repos. @claude in a pull request comment reviews code, implements changes, or answers questions. Assign it a GitHub issue and it opens a PR with an implementation.
On my projects, I open a PR and @claude catches an edge case inline. I file an issue, Claude opens a PR with a patch. Nights and weekends with limited time — this is a force multiplier.
You can run multiple instances simultaneously. One agent refactors a module while another writes tests. For enterprises — especially financial services, where strict system separation is the norm — parallel agents mirror how teams already think about service boundaries. Spinning up a new “developer” takes seconds, not a recruiting cycle.
The Claude Code SDK lets you build custom agents with the same tools. Pipe logs into Claude Code, chain it with your build system, embed it in CI.
49% to 72.7% in Seven Months Link to heading
Opus 4 works continuously for hours on complex, multi-step tasks. Not minutes. Hours.
SWE-bench scores went from 49% with Claude 3.5 Sonnet to 72.7% with Sonnet 4. A 48% improvement in autonomous coding capability in roughly seven months. The trajectory is steep and shows no sign of flattening.
Twenty years of writing software, and the tools got incrementally better every year. Better editors, better debuggers, better frameworks. This isn’t incremental. This is a phase change.
What Comes Next Link to heading
I’m all in on Anthropic. Nights and weekends learning their tools, building on their platform, rethinking how I architect systems. Time is money, and I’ve chosen to invest mine here.
Anthropic doesn’t just ship models — they define the standards the industry converges on:
- Model Context Protocol — Anthropic created this open standard for AI-tool integration in November 2024. By May 2025, OpenAI adopted it, Google DeepMind confirmed support, and Microsoft announced Windows integration.
- Constitutional AI — RLAIF that replaced brute-force human labeling with principled self-alignment. Published December 2022, now referenced across the industry.
- Responsible Scaling Policy — The ASL framework for measuring and mitigating AI risk, September 2023. Other frontier labs subsequently adopted similar approaches.
- Mechanistic Interpretability — Pioneered techniques to understand what’s happening inside neural networks. Foundational research others have built upon.
- Computer Use — First frontier AI model to ship autonomous desktop control, October 2024.
Claude Code is where this converges into a developer tool. GitHub Actions integration is live. VS Code and JetBrains extensions shipped in beta alongside Claude 4. The SDK opens the door to custom agents. All built on MCP — the standard Anthropic created and the industry adopted.
Teams that adopt agentic coding aren’t just faster. They’re operating with fundamentally different economics. The cost of experimentation drops. The cost of maintaining legacy systems drops. The barrier to building ambitious software drops.
Two days into Claude 4, and I already can’t imagine going back. The future of software development isn’t writing code faster. It’s thinking at a higher level of abstraction while an agent handles the implementation. That future arrived Thursday.