# Armin Ronacher: Agent Design Is Still Hard Source: https://www.youtube.com/watch?v=glA4YM0sX3U Recap page: https://rapidrecap.app/video/glA4YM0sX3U Generated: 2025-11-24T16:07:04.977+00:00 --- ## Quick Overview Armin Ronacher argues that agent design remains difficult due to the complexity of managing state and context, specifically highlighting that relying on agent history alone, like in basic agent loops, leads to costly failures or brittle systems; the solution involves explicit control over context, such as using a virtual file system or structured messaging, which is more reliable than simply hoping the model remembers everything. **Key Points:** - Agent design is still hard because managing state and context is complex, often leading to brittle systems or costly failures. - Relying solely on the agent's historical memory (like in simple agent loops) is unreliable and can cause the system to fail or leak internal state. - The recommended solution involves explicit context management, such as providing a virtual file system (VFS) or structured messages for the agent to use. - The author advocates for a modular approach where tools pass data back and forth using file paths, rather than relying on internal monologue or complex context injection. - The cost of complex agents (like those using Anthropic's Hiku or Sonnet) is often higher per token than simpler tools, making efficiency important. - A key takeaway is that explicit control, like using a simple command-line tool to manage state explicitly, is more robust than complex, implicit context handling. - The speaker suggests that the most effective approach for complex, ongoing agent tasks is often super-minimalist, relying on explicit state management rather than trusting the model's internal memory. ![Screenshot at 0:00: The introductory screen for the 'AI Papers Podcast' featuring two hosts/speakers overlaid on a grid with a fluctuating waveform, indicating a discussion format.](https://ss.rapidrecap.app/screens/glA4YM0sX3U/00-00-00.png) **Context:** The video features a discussion, likely from a podcast or presentation by Armin Ronacher, focusing on the inherent difficulties in designing effective, reliable AI agents. The core theme revolves around the challenge of maintaining context and state over long operational periods, especially when agents perform complex, multi-step tasks that might involve memory loss or 'hallucination' if context is not explicitly managed. ## Detailed Analysis Armin Ronacher asserts that agent design remains a difficult area, contrasting simple theoretical discussions with the gritty reality of building agents that work. He critiques the common practice of relying solely on the agent's history or context window, noting that this often leads to agents forgetting crucial information or exhibiting catastrophic failures when dealing with complex tasks. The primary issue discussed is the 'abstraction trap' where developers might abstract away necessary control, leading to agents that fail to manage context effectively. Ronacher strongly advocates for explicit control mechanisms. He suggests treating the agent's memory like a virtual file system (VFS) where tools explicitly read from and write to named files (or paths), rather than relying on the agent to simply recall information from its internal monologue or history. This approach, which he calls 'systems engineering,' ensures that the agent's state (like cash points or task status) is reliably maintained, even across different tools or runs. He contrasts this with relying on the model's inherent context, which he finds unreliable, citing an example where a model failed to recall a critical context point after being prompted seven times, leading to a catastrophic failure that lost 50,000 tokens worth of work. The conclusion is that explicit, modular control over context and state is superior for reliability, even if it seems less 'elegant' than trusting the model implicitly. ### The Abstraction Trap - Moving past white papers and theoretical discussions - The practical reality of building agents that work - Agent design is still hard ### The Core Problem - Relying only on agent history leads to brittle systems - Failures often stem from context loss or memory issues - Explicit control is necessary ### Recommended Solution - Use explicit context management like a Virtual File System (VFS) - Tools read and write to named file paths - This decouples tools and increases modularity ### Cost and Efficiency - Models like Anthropic's Hiku/Sonnet can be expensive per token - Simple command-line tools managing explicit state can be more cost-effective ### Final Takeaway - Robustness comes from explicit control over state and context, not relying on implicit memory or complex chain-of-thought prompting ![Screenshot at 0:05: The speaker begins discussing moving past white papers to the practical reality of building agents that work.](https://ss.rapidrecap.app/screens/glA4YM0sX3U/00-00-05.png) ![Screenshot at 0:24: The speaker stating that operational knowledge, not press releases, is essential for building working AI agents.](https://ss.rapidrecap.app/screens/glA4YM0sX3U/00-00-24.png) ![Screenshot at 1:37: The speaker contrasting the reliability of manual control versus automatic context handling, suggesting manual control is better for predictability.](https://ss.rapidrecap.app/screens/glA4YM0sX3U/00-01-37.png) ![Screenshot at 2:24: Visual representation of the problem: using the native SDK destroys the message history, which defeats the purpose of context.](https://ss.rapidrecap.app/screens/glA4YM0sX3U/00-02-24.png) ![Screenshot at 4:43: The speaker explaining that explicitly instructing the agent to use a file system \(VFS\) for state management is critical to avoid dead ends.](https://ss.rapidrecap.app/screens/glA4YM0sX3U/00-04-43.png)