Edition #1. Welcome to the first edition of the AI Agent newsletter, thanks for signing up.
Last night I got a push notification: Opus 4.6 is out. A few hours after the long-awaited model from Anthropic dropped, rival OpenAI came out with GPT-5.3-Codex. The timing says enough. This is a race, and both parties want to win.
Claude Opus 4.6: the details
1 million tokens of context. That means you can load a complete book, an entire codebase, or months of emails in one conversation. Previously you had to chop it into pieces. Or it simply took more time for the subagents to understand the codebase. Now that goes a lot faster.
Agent teams (swarms). Multiple AI agents collaborating on a task. One does research, another writes the code, a third reviews the code that was written. In business terms: you (the CEO) have an idea, you discuss it with your CTO (your Claude Code terminal), which then literally spawns software developers, marketers, specialists to work on your idea based on your feature requests. This changes how you work.
Adaptive thinking. The model now decides for itself how much it needs to think. For a simple question you get a quick answer. Complex problem? Then the model digs deeper, without you having to specifically ask for it.
GPT-5.3-Codex: the details
OpenAI's answer to Opus 4.6.
Best coding performance ever. 57 percent on SWE-Bench Pro, 76 percent on TerminalBench 2.0, 64 percent on OSWorld. For non-nerds: these are the standard tests with which coding AI is measured. These scores are records.
25 percent faster. And it uses half the tokens of the previous version for the same tasks. Faster and more efficient.
Live steering. You can correct the model while it's working. See it going the wrong way? Just intervene and steer. That couldn't be done before. They call this mid-task steerability.
Computer use. The model can now also operate your computer. Click, type, navigate. Still experimental, but it works.
The model that built itself. GPT-5.3-Codex is the first model that helped train and deploy itself. OpenAI's team used early versions to fix bugs, monitor training, and analyze tests. AI building AI. That's where we are now.
What does this mean for us?
Developer? The smartest devs I follow don't pick one model, they switch by task. Opus 4.6 for planning architecture and complex problems. GPT-5.3-Codex for fast execution and debugging. Not either-or, but both.
And those 1 million tokens of Opus 4.6 context? For coding 200k is often enough. Where it does perform: when you want to analyze a complete codebase in one go.
Working in an office? Here Opus 4.6 gets very interesting. 90.2 percent score on legal reasoning. Context window of 1 million tokens. Search through months of emails. Analyze full annual reports. Summarize a quarter's worth of customer feedback. Things that normally take days, you now do in minutes.
Are you an entrepreneur? This is where agent teams shine. Opus 4.6 can coordinate multiple sub-agents without losing the thread. Multi-day workflows that don't forget context. A research agent reading tons of documents and remembering everything. An assistant agent that knows your full project history. A coordinator agent driving your team of AI agents. One agent does your market research, another writes your content, a third checks everything. Without you having to coordinate. This works now.
The question isn't whether you'll use it, but when your competitors will.
Practical tips to get more out of it
Let Claude think. For complex questions it pays to ask: think this through carefully before answering.
Use the 1m context size of Opus 4.6. Ask your Agent to analyze entire documents for connections. Analyze an annual report. Summarize months of customer feedback.
Use subagents and split big tasks. For complex projects you can ask Claude to execute sub-tasks in parallel. Just add use subagents to your prompt for more compute.
Run /insights in your Claude Code terminal. Claude will then analyze and summarize your message history with Claude and suggest how you can further improve your workflow.
Claude Code GitHub commits are growing exponentially
Currently 4 percent of all GitHub commits are written by Claude Code. That's 134,000+ per day. The expectation is that this will be 20+ percent by end of 2026. Not surprising, when even the creator of Claude Code admits to no longer writing a single line of code himself.
If you've ever had an idea for a tool but thought I can't program, that barrier is now gone. It costs you at most an evening to discover what's possible.
Use your AI Agent as your daily working partner
Most people open their AI Agent, ask one question and close it again. That's like renting a Ferrari and only sitting in it once to feel what it's like.
The real value sits in making optimal use of the AI Agent. OpenClaw is an open-source AI agent that runs on your own computer or server and executes tasks via WhatsApp or Telegram. Sending emails, managing your calendar, writing code, just by sending a message. The project went viral with 145k+ GitHub stars in a few days. It's not simple. The agent gets access to all your accounts, so you need to know exactly what you're doing security-wise.
Personally I've been using OpenClaw for a month, and it's a true gamechanger. The reason I get so much out of it is because I've fully customized it with unique skills I created and software I wrote that it uses to be an even better assistant for me.
OpenClaw currently handles morning standups, project management, critical reflection as co-CEO, day-end standups and soon also administration for me. No sick days. No vacations. Always available.
It feels strange at first. Like you're cheating. After a week you wonder how you ever worked without it.
Solopreneurs are no longer a joke
All tasks that occur repeatedly and that you can describe, can largely be automated. The role of assistant is becoming obsolete. My friends look at me strangely when I say that, but it's really true.
The race is no longer about who builds the smartest chatbot. It's about who builds the best working partner.
And that race has already begun.
