# Advice on prompt-engineering with AI agents | Peter Steinberger and Lex Fridman Source: https://www.youtube.com/watch?v=BuvYFWrH_WQ Recap page: https://rapidrecap.app/video/BuvYFWrH_WQ Generated: 2026-02-15T13:32:25.677+00:00 --- ## Quick Overview The conversation between Lex Fridman and Peter Steinberger explores the 'Curve of Agentic Programming,' illustrating how the complexity of prompt engineering and agent orchestration rises sharply before leveling off as agents become more capable of handling complex tasks with simpler instructions. Steinberger notes that early interactions involve simple prompts like "ple fix this," complexity peaks with orchestration of 8 agents, custom workflows, and 18 slash commands, and eventually simplifies again to high-level instructions like "Hey, look at these files [...] and then do [med size change]." **Key Points:** - The 'Curve of Agentic Programming' shows complexity peaking when orchestrating multiple specialized agents (e.g., 8 agents at once) with complex workflows and tool usage (like 18 different slash commands). - The complexity decreases after the peak, leading to simpler, higher-level prompts like 'Hey, look at these files [...] and then do [med size change]' when agents achieve greater capability. - Steinberger shares his past reliance on writing very long, bespoke prompts, contrasting with the current trend of using voice commands for terminal operations. - The difficulty lies in the complexity of orchestration, not necessarily the individual LLM's programming ability; the system needs to be built to empathize with the agent's perspective. - A major challenge is the 'waterfall model' approach to software development, where agents are expected to handle the whole process, which often fails if the underlying agent is not good. - The speaker suggests that building empathy for the AI agent's required context is crucial for effective prompt engineering, as agents often start from 'nothing' or a bad codebase default. - The speaker mentions his personal development process involved spending an 'ungodly time' over the year playing with and iterating on his prompts and orchestration layers to improve agent performance. ![Screenshot at 00:04: The 'Curve of Agentic Programming' graph is displayed, showing complexity rising steeply from simple prompts like 'ple fix this' to a peak involving complex orchestration, before declining towards high-level instructions like 'Hey, look at these files \[...\] and then do \[med size change\]'.](https://ss.rapidrecap.app/screens/BuvYFWrH_WQ/00-00-04.jpg) **Context:** This segment features Lex Fridman interviewing Peter Steinberger, where they discuss the evolution and challenges of programming AI agents, particularly focusing on prompt engineering and the complexity involved in managing multi-agent systems. Steinberger introduces a conceptual graph, the 'Curve of Agentic Programming,' to illustrate how the required level of specificity in prompts changes over time as the underlying AI capabilities and orchestration systems mature. ## Detailed Analysis Lex Fridman and Peter Steinberger discuss the concept they term the 'Curve of Agentic Programming.' This curve plots complexity over time, illustrating how initial simple instructions like 'ple fix this' (00:05) escalate in complexity when developers attempt to chain together multiple agents (up to 8 simultaneously) using complex orchestration, multi-checkouts, custom subagent workflows, and a library of 18 different slash commands (00:04). This complexity peaks when attempting to manage all these moving parts. However, as agent capabilities improve, the complexity curve descends, allowing users to issue simpler, high-level commands, such as asking the agent to review files and perform a 'med size change' (00:05). Steinberger contrasts his early habit of writing very long prompts with the current approach of using voice commands for terminal operations, noting that voice input is often more natural (00:11-00:20). He explains that his own development process for agents involved spending a significant amount of time iterating on prompts and orchestration layers to improve agent output, suggesting that building empathy for the agent's starting context is vital (03:46-03:58). He critiques the 'waterfall model' mentality applied to agents, where one agent is expected to handle the entire software development lifecycle, noting that if the base LLM is poor, the system fails, and true skill lies in building the right architecture and empathy (03:00-03:25). ### Curve of Agentic Programming - Simple initial prompts ('ple fix this') lead to peak complexity with multi-agent orchestration (8 agents, custom workflows, 18 slash commands) before complexity drops with high-level instructions ('Hey, look at these files...'). ### Prompting Evolution - Speaker previously wrote very long prompts but now prefers voice commands for terminal operations; empathy for the agent's context is key. ### Development Challenges - The expectation that one agent can handle the entire software development lifecycle (like a waterfall model) often fails due to poor base LLM quality or insufficient empathy building. ### Iterative Improvement - Developing effective agents requires spending significant time iterating on prompts and orchestration layers, leading to better agent performance and output. ![Screenshot at 00:04: The 'Curve of Agentic Programming' graph illustrating the relationship between time and complexity in developing autonomous AI agents.](https://ss.rapidrecap.app/screens/BuvYFWrH_WQ/00-00-04.jpg) ![Screenshot at 00:03: Lex Fridman in the studio, seated opposite his guest, ready to discuss complex technical topics.](https://ss.rapidrecap.app/screens/BuvYFWrH_WQ/00-00-03.jpg) ![Screenshot at 00:14: The guest, Peter Steinberger, gesturing emphatically while discussing the challenges of agentic programming.](https://ss.rapidrecap.app/screens/BuvYFWrH_WQ/00-00-14.jpg) ![Screenshot at 02:14: Lex Fridman leans on his hand, listening intently to the guest's explanation of agentic empathy.](https://ss.rapidrecap.app/screens/BuvYFWrH_WQ/00-02-14.jpg) ![Screenshot at 03:43: Lex Fridman makes a precise gesture while describing how software engineers can overcomplicate things when interacting with LLMs.](https://ss.rapidrecap.app/screens/BuvYFWrH_WQ/00-03-43.jpg)