Agentic Engineering: The New Software Workflow
Peter Steinberger, creator of PSPDFKit and ClaudeBot, details the shift from traditional coding to agentic engineering. Learn how closing feedback loops, prioritizing architecture over syntax, and leveraging parallel AI agents redefine developer productivity and product strategy.
The Shift to Agentic Engineering
The software development paradigm is undergoing a fundamental transformation, moving from manual coding to "agentic engineering." Peter Steinberger, creator of PSPDFKit and the viral ClaudeBot, argues that the core competency for developers is no longer syntax mastery, but system architecture and prompt design. By leveraging AI agents, developers can merge hundreds of commits daily without sacrificing quality, provided they design systems that allow agents to self-verify their work.
Closing the Feedback Loop
The critical differentiator between effective and frustrating AI coding is the "closed loop." Steinberger emphasizes that AI excels at coding because code is verifiable: it can be compiled, linted, and executed. To harness this, developers must architect applications with fast, automated testing environments. When an agent can run its own tests and debug failures independently, the need for human intervention drops significantly. This approach turns the agent into a self-correcting engineer rather than a passive code generator.
The New Developer Role
The modern developer acts as a "builder" or architect, managing a team of parallel AI agents. This requires a mental shift from writing code to steering outcomes. Steinberger notes that he no longer reads most of the code he ships, focusing instead on system structure and user experience. The workflow involves parallelizing tasks across multiple agents to maintain flow state, treating the development process as a game of iterative shaping rather than linear construction. Pull requests are evolving into "prompt requests," where the quality of the initial instruction set dictates the success of the implementation.
Strategic Implications for Business
For entrepreneurs and tech leaders, this shift implies that product velocity can increase dramatically with smaller teams. Steinberger suggests that companies could operate with 30% of their previous headcount by leveraging high-agency engineers who can delegate execution to AI. However, this requires a cultural overhaul: companies must redefine roles, prioritize architectural thinking, and invest in tooling that supports agent-based workflows. The ability to iterate rapidly and cheaply allows for more experimental product development, reducing the risk of building the wrong thing. Ultimately, the value of software lies in how it feels to the user, a standard that AI agents can help achieve by freeing developers to focus on polish and experience rather than plumbing.
Key insights
-
AI agents perform best when the system allows them to close the feedback loop through automated testing and execution. This self-verification capability is the primary reason AI is effective at coding but less so at creative writing.
Impact: Enables developers to trust AI output more, reducing manual review time and increasing deployment speed.
-
The role of the developer is shifting from code writer to system architect. Success depends on the ability to define high-level structures and steer agents, rather than manually implementing features.
Impact: Redefines hiring criteria to value architectural thinking and prompt engineering over syntax proficiency.
-
Parallelizing AI agents allows developers to maintain flow state by context-switching between tasks while agents work. This significantly increases individual throughput compared to sequential coding.
Impact: Enables small teams to achieve output levels previously requiring large engineering departments.
-
CLI interfaces are more efficient for AI agents than MCPs because they allow for command chaining and filtering, reducing context window bloat. This architectural choice improves agent reliability and speed.
Impact: Guides developers to build agent-friendly interfaces, improving the overall effectiveness of AI integration.
-
Rigid upfront planning is less effective in the AI era. Iterative, low-cost prototyping allows developers to discover requirements through interaction, leading to more aligned and polished products.
Impact: Reduces time-to-market and minimizes the risk of building features that do not meet user needs.
Action items
-
Refactor codebases to include fast, automated testing loops that AI agents can execute independently. Ensure that build, lint, and test processes are optimized for speed to facilitate self-correction.
Impact: Increases the reliability of AI-generated code and reduces the need for manual debugging.
-
Train engineering teams to treat pull requests as prompt requests. Focus on the quality of the initial specification and architectural intent rather than the generated code itself.
Impact: Improves the signal-to-noise ratio in code reviews and leads to better-aligned implementations.
-
Implement parallel agent workflows where developers manage multiple AI tasks simultaneously. Use context-switching strategies to maintain flow state and maximize throughput.
Impact: Boosts individual developer productivity and allows for faster iteration on complex features.
-
Prioritize CLI-based tooling over MCPs for agent integration. Design interfaces that allow agents to filter and chain commands efficiently to minimize context usage.
Impact: Enhances agent performance and reduces the cognitive load on the AI model, leading to more accurate outputs.
-
Adopt an iterative, experimental approach to product development. Use AI to rapidly prototype and test ideas, allowing user feedback and agent interaction to shape the final product.
Impact: Accelerates product discovery and ensures that the final solution aligns with actual user needs.
Quotes
“The good thing about how to be effective with coding agent is always like you have to close the loop.”
“I feel like a human merge button because the community is like blowing off, and all I do is like reviewing PRs.”
“I would say I write better code now that I don't write code myself anymore.”