Code 2025
Agents are Robots Too: What Self-Driving Taught Me About Building Agents — Jesse Hu, Abundant
Overview
This talk draws parallels between building self-driving cars and developing AI agents, emphasizing that agents, like robots, operate in complex environments and require more than just sophisticated models. The core thesis is that the 99% of work beyond the AI model—encompassing infrastructure, tooling, feedback loops, and robust offline stacks—is crucial for reliable agent development and deployment.
Who should watch
- AI Engineers
- Product Managers
- Builders working on agentic systems
- Those interested in the practical challenges of shipping AI agents
- Individuals exploring the intersection of robotics and AI agents
Key takeaways
- The development of AI agents mirrors robotics challenges, with 99% of the effort lying in integration, deployment, and offline tooling rather than just the core model.
- Closed-loop systems, where actions are followed by feedback and recalibration, are essential for agent reliability, similar to how self-driving cars adjust to real-world conditions.
- Implicit design choices in agents, such as conversational turns, differ from the continuous, real-time sampling common in robotics, impacting responsiveness.
- Agents are evolving from stateless processes to stateful ones, requiring consideration of persistent file systems and the overall environment state, impacting evaluation and simulation.
- The imitation learning problem, where agents struggle with out-of-distribution scenarios like unexpected pop-ups, is a significant challenge inherited from robotics.
- Actions have consequences in the real world, necessitating robust simulation and evaluation frameworks to handle the complexity and messiness that predictive models alone cannot address.
- Predefined interfaces, like those found in vehicles with human controls and telemetry, simplify agent development compared to tasks requiring full desktop interaction.
- Iterative development, or hill climbing, is key, with real-world logs feeding back into simulation to ground improvements and provide deeper insights beyond simple metrics.
Notable quotes
*The model is only doing 1% of the work and 99% of the work goes into other things.*
*The winning team not just having the best model and the best online stack but having the best offline stack because that enables developers to be much faster and ship more much more reliably.*
*Actions have consequences. We're not just dealing with classification models. We're not just dealing with prediction models or sequences. We're dealing with a whole new paradigm in which you predict, you act, and then you deal with the consequences of that action and then re-evaluate everything you've done before.*
Unofficial community note. Prefer the recording for nuance.