Intel / article

LLM Powered Autonomous Agents

Source: lilianweng.github.io
Author: Lilian Weng
Date: Published Jun 23, 2023

Context

An can use the result of one action to decide what to do next. In a 2023 survey, Lilian Weng explains how a language model can work inside a loop that plans steps, calls tools and receives observations. A search result or a program's output becomes information for the next decision.

Memory supplies earlier information that would otherwise fall outside the model's immediate . These components let a system attempt tasks spanning multiple actions, but each handoff introduces possible failure: a plan can become unsuitable, a tool request can be malformed, or relevant history can be missed. The survey describes an architecture for attempting autonomous work, with reliability still dependent on the whole loop.

Key takeaways · AI-distilled

Planning is two separate moves, not one. The agent decomposes a large task into manageable subgoals up front, and separately reflects on finished actions, criticizing its own mistakes and refining the next steps.
Agent memory maps onto the model itself. Short-term memory is just in-context learning, whatever you fit in the prompt. Long-term memory is an external vector store the agent writes to and retrieves from, so it can hold effectively unlimited history.
exists because the weights are frozen after and hard to change. Calling external APIs is how an agent reaches anything the weights lack: current information, code execution, and proprietary data sources.

Terms in this piece · Glossary

AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
tool use — A model's ability to call external functions — run code, search the web, edit files — instead of only generating text.
pretraining — The first, biggest phase of building a model: training it on enormous amounts of text so it learns language, facts, and reasoning in general.

Key quotes

“The potentiality of LLM extends beyond generating well-written copies, stories, essays and programs; it can be framed as a powerful general problem solver.”

“In a LLM-powered autonomous agent system, LLM functions as the agent’s brain, complemented by several key components:”

Read the source lilianweng.github.io