Vibeleaderboard
← All Intel
Intel / video

Best of 2024 in Agents (from #1 on SWE-Bench Full, Prof. Graham Neubig of OpenHands/AllHands)

Source
youtube.com
Author
Latent Space
Date

Context

Graham Neubig argues that a can work better with one general that runs code than with a long menu of narrow tools. Neubig is a Carnegie Mellon professor and maintainer of the open-source OpenHands agent framework, and made the case in a Latent Space talk on agents in 2024, calling the approach instrumental to the team's success.

Neubig's example is finding the most cost-effective of four countries to buy a phone. An agent with one tool per lookup would need about 30 tool calls. In OpenHands the agent instead writes a Python that calls the same tools as functions and runs everything at once. If the program errors, the agent sees the error and fixes it.

OpenHands gives its agent only five or six core tools: running commands and notebook cells, two ways of editing files, and web browsing. Anything else comes from Python libraries the agent can use, as a human programmer would. This is the developer's account of the design, backed in the talk by demos on Neubig's own tasks.

Terms in this piece · Glossary
  • Agent mode — A mode in which one agent can inspect, edit, run commands, and iterate toward an outcome.
  • Tool — An operation that the software around a model makes available for the model to request, with a name, expected inputs and a kind of result.
  • Program — A set of instructions stored so that a computer can run it.
  • Terminal and shell — A text interface for navigating files and running programs.
Read the source www.youtube.com
More from Latent Space
Recommended reads
Comments

Checking sign-in…

Loading comments…