World's Fair 2024
Building Reliable Agentic Systems: Eno Reyes
Overview
This talk explores practical lessons learned from building autonomous software engineering systems, referred to as droids. It defines agentic systems by three core characteristics: planning, decision-making, and environmental grounding. The discussion emphasizes strategies for enhancing reliability and effectiveness in these systems, drawing inspiration from control systems, robotics, and symbolic AI.
Who should watch
- AI Engineers
- Product Managers
- Builders of autonomous systems
- Those facing challenges in agentic system reliability
- Developers working with code generation, documentation, or testing agents
Key takeaways
- Agentic systems are characterized by planning (sequencing future actions), decision-making (reasoning with criteria), and environmental grounding (interacting with external systems).
- Planning can be improved using techniques like pseudo common filters for consistent reasoning and subtask decomposition for finer control, though increased decomposition risks overwhelming the agent.
- Model predictive control enables adaptive replanning based on real-time feedback, crucial in dynamic environments or when humans are involved.
- Explicit plan criteria, such as instruction tuning or hard-coded logic, can significantly improve downstream performance, especially for domain-specific challenges.
- Decision-making can be enhanced through consensus mechanisms (e.g., self-consistency, prompt ensembles) for higher accuracy at increased computational cost.
- Explicit and analogical reasoning, like Chain of Thought, helps reduce decision complexity by outlining criteria, though broad domains or large action spaces present challenges.
- Environmental grounding involves building custom AI computer interfaces for workflows not covered by existing tools, particularly effective in domains with numerous development tools like software engineering.
- Designing explicit feedback processing, such as parsing logs or enabling self-reflection on error trajectories, is critical for grounding agents and improving their self-criticism.
Notable quotes
*Agentic systems have a lot of different interpretations, a lot of different definitions.*
*The existence of an agent within an external environment is very critical to understanding some of the unique properties of agentic systems when you're actually implementing them.*
*Ultimately, all of this is about error reduction and keeping your trajectories as successful as possible as long as possible.*
Unofficial community note. Prefer the recording for nuance.