An AI agent is software that uses an AI model to pursue a goal through actions. Rather than returning a single answer, it can select tools, gather information, interpret results, and continue through a multistep operation. In one common setup, an LLM controls a virtual browser-like search workflow; in reinforcement learning, an agent instead learns decisions from rewards and penalties. section 7.2 section 10.9
A practical agent workflow can include:
- Giving the model tools, such as web search and math utilities. section 7.2
- Turning a prompt into a plan, searching for relevant evidence, and using observations to choose follow-up actions. section 7.2
- Exposing progress notes so a person can inspect the agent’s thoughts, actions, and observations rather than treating the result as magic. section 7.2
- Learning a policy from feedback when the task is reinforcement learning: act in an environment, receive rewards or penalties, and adjust strategy. section 10.9
Research agent with search tools
A LangChain agent can receive a question about an author’s technology-related titles, search the web for the author and their books, then use the findings to infer the most common topic. The model is configured with temperature 0 for accuracy, while SerpAPI supplies search access. It is a useful automation pattern, but its intermediate steps and final conclusion still deserve inspection. section 7.2
|
|
Entering new AgentExecutor chain… I should search for information about David Clinton’s books. Action: Search Action Input: “David Clinton books” Observation: DAVID CLINTON is a Linux server administrator with experience working in IT infrastructure in academic and enterprise environments. … Thought: I should look for more specific information about the books he has published. Action: Search Action Input: “David Clinton books published” Observation: Top David Clinton titles … The Complete LPI Web Development Essentials Exam Study Guide: Learn the basics of HTML, CSS, JavaScript, Express.js, and Node.js and … Thought: I should look for reviews of his books to determine the most common topic. Action: Search Action Input: “David Clinton books reviews” Observation: Books by David Clinton (Author of AWS Certified Solutions Architect Study Guide with 900 Practice Test Questions). Avg rating: 4.06 1,161 ratings 76 reviews. Thought: I now know the most common topic of David Clinton’s books. Final Answer: The most common topic of David Clinton’s books is web development and cloud computing. Finished chain. “The most common topic of David Clinton’s books is web development and cloud computing.” |
The execution trace makes the loop visible: the agent proposes a search, receives an observation, notices what remains unknown, and launches a more specific search. section 7.2
Autonomous code-review loop
Two generative models can be arranged so one produces application code and the other critiques it; the first revises repeatedly. Fully autonomous repetition is risky: a model can give flawed feedback or misinterpret it, even deleting working code, while API-token costs accumulate. section 11.25
Agents can extend AI into research, productivity work, and data analysis, especially when integrations provide live information. More autonomy is not automatically better: tool outputs, model feedback, and repeated iterations can all introduce mistakes or wasteful cost. section 7.13 section 11.25
