Peter Steinberger built PSPDFKit, which runs on over a billion devices — and now he openly admits he ships AI-generated code he doesn't even read. That got my attention. The video covers his approach to "closing the loop" with AI agents, and I've pulled out the key takeaways below.
The creator of Clawd: "I ship code I don't read"
I've been watching developers work for three decades. Last month, something clicked that I can't stop thinking about.
Peter Steinberger—the guy who built PSPDFKit, a PDF framework running on over 1 billion devices—said something in an interview that made me sit up straight: "I ship code I don't read."
Not "I ship code I didn't write." That's been true for years with copy-paste from Stack Overflow. He said he doesn't *read* it. The AI writes it, the AI tests it, the AI fixes it. He reviews the outcome, not the implementation. His open-source project Clawdbot (now called OpenClaw) hit 30,000 GitHub stars within weeks of launching in January 2026. Clearly, other developers are paying attention.
Here's what this means for you, even if you've never written a line of code: the cost and timeline of building custom software is about to collapse. And the businesses that understand this shift will have an enormous advantage over those who don't.
In a minute, I'll show you the real bottleneck that's emerging—and it's not what you'd expect.
What Are AI Coding Agents (And Why Should You Care)?
You've probably heard of tools like GitHub Copilot or ChatGPT helping developers write code. Those are AI *assistants*. They wait for a developer to ask a question, suggest some code, and then wait again. The developer is still in the driver's seat, making every decision.
AI coding agents are different. They're autonomous software that can plan, execute, and complete multi-step coding tasks with minimal human intervention. Think of the difference between a calculator and an accountant. A calculator does what you tell it. An accountant understands your goal—"minimize my taxes legally"—and figures out the steps to get there.
Here's a simple mental model: an AI coding agent works in a loop. It understands the goal. Decides the next step. Uses a tool or asks a question. Checks the result. Repeats until done—or escalates when it's stuck.
While 92% of enterprises are increasing their AI investment, the competitive advantage isn't coming from tools that generate text or suggest code snippets. It's coming from tools that automate execution—that can take a goal and work toward it independently.
How AI Coding Agents Actually Work
The secret to making AI coding agents effective isn't fancy technology. According to Steinberger, it's "closing the loop"—making sure the agent can debug and test itself.
Here's what that means in practice: when you give an AI coding agent a task, you don't just ask it to write code. You make sure it can:
- Write the code
- Write tests for that code
- Run those tests itself
- See what failed
- Fix the failures
- Repeat until the tests pass
This is why Steinberger can ship code he doesn't read. The agent isn't just generating text—it's validating its own work. If the tests pass and the feature works as specified, does it matter exactly how the code is structured?
Even for Mac apps—traditionally complex to build—he has the agent create a command-line debugging tool that exercises all the same code paths. Then the agent can iterate and fix issues without human intervention.
The Part Most Business Owners Miss About AI Coding Agents
Here's the bottleneck shift I promised to explain—and why it matters more than the technology itself.
For decades, the limiting factor in software development was coding capacity. You had more ideas than your team could build. Hiring was hard. Projects took months. Every feature had to fight for developer time.
That's flipping. Organizations are shifting from coding bottlenecks to idea generation bottlenecks. When an AI agent can take a well-specified task and execute it in hours instead of weeks, suddenly the constraint isn't "can we build it?" but "should we build it?" and "how exactly should it work?"
Some teams are even moving toward what's being called "agent-native architectures"—where prompts to agents define product features rather than detailed code instructions. The code becomes an implementation detail. The specification becomes the product.
This is why Steinberger suggests that pull requests might become "prompt requests." Instead of reviewing code line-by-line, you review the instructions that generated the code. If the tests pass and the behavior is correct, the implementation is secondary.
For business owners, this means: your ability to clearly specify what you want is becoming more valuable than your ability to pay for developer hours.
What This Means for Agentic Engineering in Your Business
Even if your business doesn't have a development team, this shift affects you. Here's how:
**Custom software becomes affordable.** That internal tool you've always wanted—the one that would save your team 10 hours a week but wasn't worth a $50,000 development project? It might now cost a fraction of that. The agentic AI market is projected to grow at 45% annually and reach $47 billion by 2030 because businesses are discovering they can automate processes that were previously too expensive to address.
**Speed advantages compound.** If your competitor can iterate on their software twice as fast, they can test ideas twice as fast. Over a year, that compounds into a significant gap. The global financing gap for small and midsize enterprises is $5 trillion—and agentic AI is helping to close it by making sophisticated tools accessible to businesses that couldn't previously afford them.
**Integration becomes practical.** AI agents for business are increasingly able to connect with messaging platforms, business tools, and databases. OpenClaw (the project that went from 0 to 64,000 GitHub stars in days) integrates with over 50 platforms including WhatsApp, Telegram, Slack, and Discord. This kind of flexibility used to require expensive custom development.
Where AI Coding Agents Fall Short
I'd be doing you a disservice if I didn't cover the downsides. There are real limitations.
**The mental load problem.** Steinberger admits that working with multiple AI coding agents is mentally more exhausting than traditional coding. Instead of managing one task deeply, you're context-switching between five or ten parallel agents. "I don't have one employee that I manage," he says. "I have like five or ten that all work on things and I switch from this one part to this other part."
**The fundamentals still matter.** Here's what nobody selling AI tools will tell you: the fundamentals of effective software development haven't changed. Good specs. Clear documentation. Proper reviews. The right technology stack. A history of decisions and why they were made. AI agents don't eliminate the need for these—they amplify the cost of not having them.
**Quality control gets harder, not easier.** When code gets generated faster, the temptation is to skip thorough testing. But the bugs don't disappear—they just arrive faster. You still need humans who understand what the software should do and can verify it actually does that.
How to Evaluate AI Coding Agents for Your Team
Whether you're managing developers or hiring contractors who use these tools, here's what to look for:
- **Can it close the loop?** The agent should be able to write tests and run them itself. If it can only generate code and wait for human testing, you're getting an assistant, not an agent.
- **What's the escalation path?** Good agents know when they're stuck and ask for help instead of confidently producing garbage. Ask how the tool handles uncertainty.
- **How does it handle existing code?** Generating new code is easy. Understanding and modifying an existing codebase is hard. Test this specifically.
- **What's the debugging story?** When something breaks—and it will—how does the development team figure out what went wrong? "The AI did it" isn't an acceptable answer.
The key question to ask any vendor or developer: "Show me a task that failed and how you diagnosed and fixed it." Anyone can show you success stories. The failures reveal the real capability.
Your First Week With AI Coding Agents
If you want to understand how this technology could affect your business, here's a practical starting point:
- **Identify one repetitive software task.** Something your team does manually that could be automated. Start small—a report that gets generated weekly, a data transformation, an internal tool.
- **Write a clear specification.** Describe what the task should accomplish, not how to code it. Include what success looks like. This exercise alone will reveal how clearly you understand your own processes.
- **If you have developers:** Ask them to try completing this task using an AI coding agent with your spec. Compare the time spent writing the spec + agent iteration vs. traditional development.
- **If you don't have developers:** Use this spec to get quotes from contractors. Ask specifically whether they use AI coding agents and how that affects their timeline and pricing.
- **Measure the outcome, not the method.** Does the solution work? Does it pass the tests? Does it solve the business problem? If yes, the implementation details matter less than you think.
- **Budget 4-8 hours for the experiment.** That's enough time to write a solid spec (2 hours), iterate with an agent or contractor (4-6 hours), and evaluate the result.
The goal isn't to adopt AI coding agents immediately. It's to understand the new economics of software development so you can make informed decisions about your technology investments.
What AI Coding Agents Mean for Your Technology Roadmap
- AI coding agents are autonomous software that can plan, write, test, and fix code with minimal human intervention—fundamentally different from AI assistants that just suggest code snippets.
- The competitive advantage is shifting from coding capacity to specification quality. Clear requirements are becoming more valuable than developer hours.
- The agentic AI market is projected to reach $47 billion by 2030 at 45% annual growth—this isn't a niche technology.
- The "close the loop" principle is essential: effective AI coding agents must be able to test their own work and fix their own mistakes.
- Mental load increases even as coding time decreases. Managing multiple parallel agents requires different skills than traditional development oversight.
- The fundamentals haven't changed: good specs, clear documentation, proper testing, and decision history are more important than ever—AI agents amplify both good and bad practices.
Frequently Asked Questions About AI Coding Agents
Do I need to understand coding to benefit from AI coding agents?
No. The shift toward "agent-native architectures" means the emphasis is moving from coding knowledge to specification clarity. Your ability to clearly describe what you want—the business outcome, the success criteria, the constraints—is becoming more valuable than technical knowledge. That said, having someone on your team who understands software development basics helps you evaluate whether the AI's output actually solves your problem.
How much do AI coding agents cost compared to traditional development?
It varies significantly, but the economics are changing fast. Tasks that might have taken 40 hours of developer time can sometimes be completed in 4-8 hours with an AI coding agent (including specification writing and iteration). However, complex projects with existing codebases or unusual requirements still require substantial human expertise. The savings are most dramatic for well-defined, greenfield tasks.
What's the difference between AI coding agents and tools like GitHub Copilot?
GitHub Copilot is an AI assistant—it suggests code when you ask, but you're still driving. AI coding agents are autonomous: you give them a goal, and they figure out the steps, write the code, test it, and iterate until it works. The distinction is like the difference between a GPS that gives you directions and a self-driving car that takes you to your destination.
Can AI coding agents work with my existing software and systems?
Modern AI agents for business can integrate with dozens of platforms. OpenClaw, for example, connects with over 50 messaging and business tools. However, integrating with legacy systems or proprietary software often requires custom development work. The simpler and more standard your existing technology stack, the easier integration will be.
What happens when an AI coding agent makes a mistake?
Good AI coding agents are designed to catch their own mistakes through automated testing—that's the "closing the loop" principle. When they can't fix something, they should escalate to a human rather than confidently producing broken code. The key is having clear success criteria and tests that can verify whether the output actually works. Without those, mistakes can slip through faster than ever.
