This paper details the unique, AI-centric software development workflow of Peter Steinberger, creator of Moltbot (formerly Clawdbot), enabling him to ship an unprecedented volume of code, exemplified by over 6,600 commits in a single month. Steinberger's methodology centers on leveraging AI agents to perform the bulk of coding tasks, fundamentally altering traditional software engineering paradigms.

The core methodology revolves around "agentic engineering," where AI agents are integrated into a self-verifying, high-speed development loop. Key aspects include:

1. Autonomous Verification (Closing the Loop): Agents are designed to independently verify their own work. This involves empowering them to perform compilation, linting, execution of code, and validation of outputs. This self-contained feedback loop is critical for enabling the rapid iteration and high trust in AI-generated code, allowing Steinberger to "ship code I don't read." The underlying principle is that the agent, after generating code, directly executes a battery of tests and checks to confirm correctness and adherence to standards.

2. Shift from Code Review to "Promot Requests": Instead of traditional pull requests (PRs) focused on reviewing human-written code, Steinberger introduces "promot requests." The primary artifact for human review becomes the prompt used to generate the code, rather than the generated code itself. This signifies a fundamental shift in the unit of development and review, emphasizing the design of clear, effective prompts as the critical intellectual endeavor. Code reviews are largely replaced by architectural discussions, where high-level system structure, extensibility, and modularity are debated, leaving implementation details to the agents.

3. Parallel Agent Execution and Strategic Planning: Steinberger manages multiple AI agents (typically 5-10) concurrently, each working on different features or tasks. This parallelization significantly amplifies development throughput. A substantial amount of time is dedicated to planning and refining tasks with the AI before execution. He engages in iterative "push-back" and refinement of plans with the agent, often preferring models like OpenAI's Codex for long-running tasks due to its ability to execute without frequent interruptions for clarifications, unlike models such as Claude Code. This involves a sophisticated meta-prompting strategy to guide the AI's problem-solving process.

4. Intentional Under-Prompting: To foster innovation and explore unexpected solutions, Steinberger occasionally provides intentionally vague prompts. This allows the AI agents to leverage their inherent exploratory capabilities and potentially discover novel approaches that a human might not have initially considered.

5. Local CI for Agent-Driven Development: Recognizing the latency of remote Continuous Integration (CI) pipelines (e.g., 10+ minutes), Steinberger opts for local CI directly executed by the agents. This allows for immediate feedback and iteration within the agent's workflow, bypassing external dependencies and accelerating the development cycle. The agents execute tests and checks on the developer's local machine, ensuring rapid validation.

6. Emphasis on System Design over Implementation Details: Steinberger posits that a significant portion of application code involves "boring data transformation." His workflow delegates these repetitive, mechanistic coding tasks to AI agents, freeing the human engineer to focus on higher-level system architecture, design, and ensuring extensibility and modularity of the overall product. This requires a strong architectural vision to act as the "benevolent dictator" guiding the project's direction. Engineers who excel in this setup are outcome-oriented, prioritizing shipping products over meticulous algorithmic optimization or manual code construction.

This methodology has allowed Moltbot to achieve unprecedented growth and popularity, demonstrating a practical application of AI agents to scale individual productivity far beyond traditional limits by changing the nature of software engineering work itself.