Agentic Engineering: Why Peter Steinberger Stopped Reading Code

Software development is facing a paradigm shift that goes far beyond simply using AI assistants. Leading this movement is Peter Steinberger, founder of PSPDFKit and creator of the viral AI agent OpenClaw (formerly ClawdBot). His radical thesis: programmers should stop reading code.

Steinberger draws a strict line between the often-ridiculed “Vibe Coding” - chaotic late-night hacking - and professional “Agentic Engineering.” The latter is a disciplined practice where the developer is no longer the craftsman laying every brick, but the architect directing an army of AI agents.

Here are the core pillars of this new philosophy, as Steinberger demonstrates through the OpenClaw project.

”Ship Without Reading”: Trust Through Verification

Perhaps the most controversial practice in Steinberger’s workflow is shipping code to production without reviewing it line by line. To traditional developers, this sounds reckless, but in “Agentic Engineering,” human oversight is redirected:

• Closing the Loop: Instead of checking syntax, the developer defines verification loops. The agent doesn’t just write the code - it also writes the test, runs it, and reads the error message.
• Outcome-Oriented: When the agent writes code that compiles and all tests pass, the exact syntactic path is often irrelevant. A large part of software development is “plumbing” - moving data from A to B. Steinberger’s approach frees the human from this machinery to focus on system architecture and “taste.”

The Death of the Pull Request: “Prompt Requests” Take Over

In a world where machines generate code, the classic Pull Request (PR) loses significance. A PR only shows the result (the diff), not the intent.

• Steinberger argues that prompts are more important than code. A “Prompt Request” reveals how a problem was solved and what intention lay behind it.
• In the OpenClaw community, contributions are often no longer evaluated based on the code itself, but on the prompt that generated it. This allows maintainers to reproduce or adapt solutions using their own agents, rather than manually debugging someone else’s code.

CLI Over MCP: Against Context Pollution

While the industry tries to establish complex standards like the Model Context Protocol (MCP), Steinberger relies on proven Unix philosophy: Command Line Interfaces (CLIs).

• The Problem with MCPs: They tend to “pollute” the AI model’s context with unnecessary data (e.g., a weather API returns all metric data when all that was needed was “Rain: Yes/No”).
• The CLI Solution: Agents are trained to operate terminals. Using tools like jq, they can filter CLI output before it enters their context. This keeps the model’s “memory” clean and efficient. OpenClaw uses an “army of CLIs” to control everything from calendars to home automation.

OpenClaw: The Agent with a “Soul”

The OpenClaw project (formerly known by various names like Clawdbot or Moltbot, until trademark issues with Anthropic forced a rename) is proof that these ideas work in practice.

• Self-Modifying Code: OpenClaw is self-aware. The agent can read its own source code, diagnose errors, and update itself.
• The soul.md: To give the agent personality, Steinberger uses Markdown files (soul.md, identity.md) that define core values and character traits. The agent adapts these files based on interactions, simulating a persistent “relationship” with the user.
• The Operating System of the Future: OpenClaw points to a future where apps are degraded to mere APIs. Why open a fitness app when the agent knows where you are and can process the data directly? Steinberger predicts that 80% of today’s apps could be replaced by such proactive agents.

Conclusion: From Coder to Builder

Peter Steinberger’s journey shows that the role of the software developer is not disappearing - it is radically transforming. The ability to memorize syntax is being replaced by the ability to design systems, orchestrate agents, and - most importantly - contribute taste and vision. Those who still manually review every line of code today may soon be outpaced by those who have learned to build at inference speed.

Sources: Based on Peter Steinberger’s blog post “Shipping at Inference-Speed” and interviews on “The Pragmatic Engineer” and the Lex Fridman Podcast.