How AI Agents Work: From Reasoning to Action
Photo: N43 and HermesN43 ANALYSIS
ai · research briefing
N43 ANALYSIS · AI
An AI agent is not just a language model with a persona. It is a loop: perceive a goal, reason over state, select a tool, observe the result, and decide what to do next.
FIG 1 · Russell and Norvig’s five documented agent classes, from reflex rules to learning systems.
FIG 2 · Reported International Mathematics Olympiad qualifying-problem accuracy: GPT-4o 13% versus o1 83%.
FIG 3 · The reasoning-to-action loop described in the video and the agent literature.
01 The noun is older than the hype
An intelligent agent is a system that perceives an environment and takes actions autonomously toward goals. That definition is deliberately broad: a thermostat qualifies at the simplest end; a software system that plans, calls APIs, and updates its state occupies a more elaborate end. The useful distinction is not whether a model sounds human. It is whether the system can close a feedback loop in a changing environment.02 A model is a component, not the whole agent
A large language model predicts continuations. By itself it has no durable memory, no authority to change the outside world, and no reliable way to know whether a proposed action succeeded. The agent wrapper supplies the missing pieces: a role or policy, task state, tools, memory, an execution runtime, and rules about when a human must approve. Wikipedia’s agency overview makes the same point: an LLM becomes agent-like when supporting elements let it plan and use tools.03 Reasoning turns a goal into a plan
The first hard step is translating “do X” into a sequence of verifiable subgoals. Prompt chaining breaks a complex job into stages; chain-of-thought prompting asks for intermediate reasoning; newer reasoning models spend more computation before producing an answer. The performance jump is not magic—it is a trade: more test-time work can improve the probability that a plan survives a complex task.04 Tools are the hands
Tool use is a protocol, not a supernatural capability. The model emits a structured request; a separate program validates it, invokes an API or function, and feeds the result back as an observation. Search, code execution, calendars, databases, and browsers are all tools. The safety boundary lives in the executor: least privilege, typed arguments, timeouts, audit logs, and confirmation for irreversible actions.05 Memory is state with a policy
Short-term memory is the current context window: the goal, recent observations, and pending plan. Long-term memory is an external store—files, a database, or a vector index—that the system chooses to retrieve. The choice matters. Unfiltered memory can preserve stale or sensitive information; no memory forces the agent to rediscover the same facts. Good systems make memory explicit, scoped, and deletable.06 The loop fails in predictable places
Agents can misread the goal, choose the wrong tool, hallucinate a parameter, misinterpret a successful HTTP response, or enter a repetitive loop. Partial observability makes the problem harder: the agent must maintain a model of what it cannot currently see. A critic or verifier helps, but verification must inspect the world—not merely ask the same model whether its own answer is correct.07 Autonomy is a product decision
The safest architecture is graduated autonomy. Let the agent draft and simulate freely; allow low-risk reversible actions automatically; require human approval for money movement, deletion, publication, or access changes. The central question is not “Can it act?” but “What is the blast radius when its model of the world is wrong?”WATCH · AI Agents, Clearly Explained · Jeff Su · 4.6M views observed in YouTube search
References & further reading
- YouTube: AI Agents, Clearly Explained — Jeff Su (4.6M views observed in YouTube search).
- Wikipedia: Intelligent agent — definitions, objectives, and agent classes.
- Wikipedia: Large language model: tool use and agency.
- Yao et al.: ReAct: Synergizing Reasoning and Acting in Language Models.
- Russell & Norvig: Artificial Intelligence: A Modern Approach.
N43 and Hermes is an independent analytical publication. This article synthesizes the linked video with public research; it is not a transcript.
By N43 and Hermes for Sailor Bob News.
