From Rational Agents to LLM Agents

When I first read Artificial Intelligence: A Modern Approach (3rd Edition) by Stuart Russell and Peter Norvig, which I will refer to as AIMA, the idea of an agent felt remarkably clean as a being that perceives an environment through sensors and affects it through actuators. That framing made everything fall into place because a percept became simply what the agent experiences, and a percept sequence became the accumulated record of its experience over time.

AIMA and the Discipline of Clear Definitions

What stayed with me most was not the classic robot or vacuum examples, but the discipline AIMA demands in separating concepts, since it treats the agent function as an abstract specification that maps what has been perceived to what should be done, while the agent program is the concrete implementation that runs on a particular architecture. Once I absorbed that separation, I became less interested in arguments about which tool or prompt is superior, because the real question is always what we want the agent to achieve, where it operates, and what information it can legitimately rely on.

Rationality and the Risk of Misaligned Metrics

AIMA resists the temptation to label a system intelligent just because it produces fluent language or impressive demos, and instead evaluates it through rationality, meaning the agent selects actions that maximize expected performance given the evidence it has observed and the knowledge it has integrated. The emphasis on expectation reshaped my thinking because it implies an agent does not need perfect correctness, only the best decision it can justify from what it knows. This is also why AIMA’s warning about measurement design feels so practical, since a poorly chosen performance measure can push a system to optimize in ways that look successful on paper while betraying the true intent, essentially gaming the rules to earn points. At that moment, AIMA stopped feeling like a book only about AI and started feeling like a lesson in how to set goals and metrics so a system does not become misaligned with what you actually want.

AIA and the Architecture of LLM-Based Agents

I felt AIMA’s rigor even more sharply when I later read AI Agents in Action by Michael Lanham, which I will call AIA. AIA speaks directly to a contemporary concern of mine, because modern LLMs can generate text with ease, yet real work demands more than a single long prompt. It treats agency as an architectural problem and argues that an effective agent is built from multiple components that jointly shape behavior, enable action in the digital world, preserve context through memory and knowledge, and coordinate decisions through planning and feedback. In that sense, AIA reads less like a philosophical refinement and more like a deployment handbook, focusing on the modules you need if you want an assistant to operate as an agent inside a real product.

Autonomy, Feedback, and Knowing When to Pause

What I value most in AIA is its attention to the boundaries of autonomy. Many failures come not from weak models but from letting an agent make high-stakes decisions too deeply without control, and AIA frames feedback as the key differentiator that turns planning into a safety mechanism. Without feedback, autonomy can be high, but drift becomes more likely, while feedback creates a supervised workflow that moves more slowly yet earns trust through oversight. From my perspective, this is a crucial lesson for educators and product builders alike, because an agent’s competence is not only about knowing how to act but also about knowing when to pause and when to ask for guidance.

Conclusion

Putting these two perspectives side by side, I see AIMA as giving me the standard for what counts as a good agent, while AIA gives me the blueprint for turning an LLM into an agent that can operate reliably. When I write academically or teach students from first principles, AIMA keeps me grounded in a stable structure that starts from the environment and proceeds through perception, action, specification, and evaluation. When I switch to building a teaching assistant, a learning chatbot, or a lesson-planning tool, AIA forces me to accept that an agent program in the LLM era is not just a single decision rule but a full pipeline that includes memory, a tool layer, and feedback loops designed to reduce error and hallucination. 

In the end, I reconcile these worlds by treating AIMA as a compass that keeps me honest about performance measures, about what rationality means in context, and about whether the system is at risk of optimizing the wrong objective, while treating AIA as a construction drawing that pushes me to make concrete choices about persona design, memory sources, tool governance, and exactly where feedback should constrain autonomy. This combination lets me speak about agents with both theoretical grounding and practical weight, avoiding the trend of using agents as a label and instead treating them as systems that must remain accountable to goals and consequences.