ReAct: Synergizing Reasoning and Acting in Language Models
by Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, Yuan Cao
Interleaves reasoning traces with task-specific actions so a model can query external sources mid-reasoning, reducing hallucination on question answering and improving success on interactive benchmarks like ALFWorld and WebShop. The pattern most agent frameworks build on.
More resources on AI Agents
Multi Agent Systems
Builds teams of cooperating agents with the crewAI framework, assigning roles, tools, memory, and task decomposition. Worked examples cover resume tailoring, technical article writing, customer support, outreach, event planning, and financial analysis.
Agent Design Patterns
Covers core agentic patterns (reflection, tool use, planning, and multi-agent group chat) implemented in Microsoft's AutoGen framework. Projects include a two-agent conversation, a reflective blog writer, a chess-playing agent, and coding agents for financial analysis.
Multi-Agent Use
Advanced crewAI course focused on shipping agent systems: external integrations, coordinating multiple models in one crew, testing with human feedback, and deployment. Projects include project planning, a Trello progress reporter, a sales pipeline, and support analytics.
Computer Use with Anthropic
Progresses from the Claude API through multimodal prompting, prompt caching, and tool calling to Anthropic's computer use feature, ending with an agent that reads screenshots and operates a desktop interface to complete tasks.
Evaluating AI Agents
Treats agent evaluation as its own discipline: adding tracing and observability, choosing between code-based checks, LLM-as-judge, and human review, then scoring router decisions, individual skills, and full trajectories through structured experiments.
Building Browser Agents
Explains how web agents perceive pages through visual and DOM structure, then plan actions like scraping, summarizing, and form filling. Also covers AgentQ, which combines Monte Carlo tree search, self-critique, and DPO for self-correcting agents.