Appendix · Autonomy

Back to core

Appendix: Index · Prev: Embodiment · Next: Safety & alignment

Overview#copy

Autonomy involves several interacting dimensions: permissions, horizon, state, delegation, reversibility, and oversight capacity. Section 04 defines the agency envelope used in this paper. The sources below explore different forms of delegated action and their operating constraints.

1) Agents for software and web tasks#copy

Key papers#copy

• WebArena: A Realistic Web Environment for Building Autonomous Agents (Zhou et al., 2023)
• SWE-bench (Jimenez et al., 2024)

Code repositories / projects#copy

• web-arena-x/webarena
• princeton-nlp/SWE-bench

2) Learning-by-exploration agents#copy

Key papers#copy

• Voyager: An Open-Ended Embodied Agent with LLMs (Wang et al., 2023)

Code repositories / projects#copy

• MineDojo/Voyager

3) Orchestration and guardrails#copy

Key papers#copy

• Constitutional AI (Bai et al., 2022)

Code repositories / projects#copy

• langchain-ai/langgraph
• microsoft/autogen