Berkeley LLM Agent Lecture 06: SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering
Paper Reading: SWE-agent & OpenHands
Paper 1. SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering
This paper investigates how interface design affects the performance of LM agents. SWE-agent is a system that facilitate LM agents to autonomously use computers to solve software engineering tasks.
ACI: Agent-computer interface, a well-designed ACI should help the LM agent understand the state of the application given previous changes
The significance of ACI design on impact of LM agent performance: compact and efficient file editing is critical to performance,
Paper 2. OpenHands: An Open Platform for AI Software Developers as Generalist Agents
OpenHands is a platform for the development of powerful and flexible AI agents that interact with the world in similar ways to those of human developer.
Dataset Leakage
LLM-based Localization
Planning and Error Recovery
Safety Mitigation
Current Challenges: code LLMs, editing, localization, planning, safety
Future directions:
Agentic Training methods
Human-in-the-loop
Broader software tasks than coding