Building with agents 0/15

Building with agents

The agent loop

An agent is a model called in a loop: it reads its context, chooses a tool call, sees the result and repeats until it decides to stop. Coding agents and agent products are variations on this loop.

Reading

WhiteBox lessons2

Papers and essays3

  1. ReAct: Synergizing Reasoning and Acting in Language Models(opens in a new tab)

    Yao et al. 2023 ICLR 2023 45 min

    ReAct: prompt a model to alternate reasoning with actions such as Wikipedia searches and to read each result before its next step, which reduced hallucination in question answering and beat imitation and reinforcement learning agents on interactive tasks.

  2. Building effective agents(opens in a new tab)

    Schluntz and Zhang 2024 Essay 25 min

    Anthropic's guide to LLM systems that separates workflows, where code decides the sequence of model calls, from agents, where the model decides its next step and tool calls, and describes five workflow patterns: prompt chaining, routing, parallelization, orchestrator-workers and evaluator-optimizer.

  3. LLM Powered Autonomous Agents(opens in a new tab)

    Weng 2023 Essay 45 min

    A survey of agent design built around three parts: planning (task decomposition and self-reflection), memory (the context window plus an external vector store) and tool use, with summaries of ReAct, Reflexion, Toolformer, HuggingGPT and the generative agents simulation.

Videos1

Building with LLMs Various speakers 1 video

People to follow1

  • Lilian Weng

    Lil'Log

    Long survey posts that explain one research topic at a time, such as hallucination in LLMs or diffusion models, with references throughout.