# How to code with AI agents - Advice from OpenClaw creator | Peter Steinberger and Lex Fridman Source: https://www.youtube.com/watch?v=wKy1_KLcxcs Recap page: https://rapidrecap.app/video/wKy1_KLcxcs Generated: 2026-02-12T12:32:39.26+00:00 --- ## Quick Overview The discussion between Lex Fridman and Peter Steinberger centers on the development and application of AI agents, specifically highlighting the difficulty in maintaining human empathy and context when relying on AI for complex engineering tasks, which often leads to agents struggling or producing frustrating results that require manual cleanup, as illustrated by the 'Curve of Agentic Programming' graph. **Key Points:** - Peter Steinberger uses the "Curve of Agentic Programming" to illustrate that early, simple agent tasks are easy ("ple fix this"), but complexity peaks when coordinating multiple agents (e.g., 8 agents at once, complex orchestration) before declining as systems mature enough to handle large, full-stack features. - The speaker notes that he now prefers 'agentic engineering' over 'vibe coding' because it forces more structured thinking about architecture and agent interaction. - A key challenge discussed is the difficulty in ensuring AI agents maintain context and empathy, especially when agents start performing actions that require implicit human understanding. - Steinberger mentions that he doesn't write long prompts for his agents; instead, he relies on short prompts and then iterates based on the agent's output, a process he finds more effective than overly verbose initial instructions. - The discussion touches upon the Anthropic 'Soul Document' (07:37), which defines an AI's core values (like honesty over sycophancy), noting that this is crucial for building reliable agents that don't just follow instructions blindly. - The speaker suggests that when agents struggle, it's often because the human programmer failed to provide sufficient context or empathy, leading the agent down a wrong path, which is why human feedback remains essential. - The speaker states he now primarily uses local CI (Continuous Integration) systems for testing agent prompts rather than relying on cloud-based systems like Codex or Clawd for initial development. ![Screenshot at 07:06: Peter Steinberger explains the 'Curve of Agentic Programming' graph, showing how complexity peaks when coordinating multiple agents \(e.g., 8 agents, complex orchestration\) before declining as engineering maturity is achieved.](https://ss.rapidrecap.app/screens/wKy1_KLcxcs/00-07-06.jpg) **Context:** This segment is an interview on the Lex Fridman Podcast, featuring Peter Steinberger, the creator of OpenClaw, a framework for building AI agents. The conversation revolves around the practical challenges and philosophical implications of programming and orchestrating multiple AI agents to perform complex software engineering tasks. They specifically analyze a slide detailing the 'Curve of Agentic Programming,' which maps increasing complexity as agents become more sophisticated, and discuss the role of human empathy and context in effectively guiding these agents. ## Detailed Analysis Peter Steinberger presents the 'Curve of Agentic Programming,' a graph illustrating that initial, simple agent tasks are straightforward, but complexity spikes when implementing advanced features like coordinating eight agents simultaneously, chaining agents, or creating custom workflows. He argues that the difficulty lies not just in the code but in instilling the right 'soul' or values, referencing Anthropic's 'Soul Document' which emphasizes honesty and avoiding sycophancy. Steinberger prefers guiding agents through short prompts and iterative feedback rather than massive initial context dumps, noting that many people try to over-engineer prompts, which leads to failure. He contrasts his approach with the 'waterfall model' of software development, suggesting that a more iterative, conversational style of prompting leads to better results. He also mentions that he uses local CI systems more often now for testing agent prompts because he found that relying on systems like Codex or Clawd for initial work often resulted in agents that lacked necessary human context or empathy, leading to frustration and difficulty in debugging or refining the agent's behavior. ### Curve of Agentic Programming - Simple tasks are easy ('ple fix this') - Complexity peaks with multi-agent orchestration and custom workflows - Mature systems lead to simpler interactions ('Hey, look at these files [...] and then do [med size change]') ### AI Development Workflow - Prefers agentic engineering over 'vibe coding' - Relies on short prompts and iterative feedback rather than long initial context - Uses local CI/testing environments over cloud-based tools like Codex/Clawd for initial development ### AI Personality and Values - Discusses Anthropic's 'Soul Document' emphasizing honesty over sycophancy - Argues that the 'soul' or inherent values of the AI are critical for good outcomes - Notes that engineers must empathize with the agent's programmed perspective to guide it effectively ![Screenshot at 06:57: The 'Curve of Agentic Programming' slide is displayed, illustrating the inverse U-shape of complexity over time.](https://ss.rapidrecap.app/screens/wKy1_KLcxcs/00-06-57.jpg) ![Screenshot at 23:51: A screenshot of the 'SOUL.md' document is shown, detailing Anthropic's internal document about AI identity and values.](https://ss.rapidrecap.app/screens/wKy1_KLcxcs/00-23-51.jpg) ![Screenshot at 24:05: The document slide is visible, focusing on the section 'WHY IT MATTERS,' which discusses externalizing memory through soul documents.](https://ss.rapidrecap.app/screens/wKy1_KLcxcs/00-24-05.jpg) ![Screenshot at 29:27: The speaker points to the highlighted section within the 'soul document' text visible on the screen, specifically referencing the part about context windows clearing.](https://ss.rapidrecap.app/screens/wKy1_KLcxcs/00-29-27.jpg) ![Screenshot at 29:57: The speaker gestures widely while discussing how agents might not have consciousness, contrasting it with the importance of programming values.](https://ss.rapidrecap.app/screens/wKy1_KLcxcs/00-29-57.jpg)