Agentic AI is a system where an LLM takes autonomous actions toward a goal. The model can plan, act and reflect. A static prompt only gets 1 response, but an agent keeps working until it reaches the goal.

What an agent combines:

  • reasoning (CoT) with decision-making and tool use
  • memory, so that it can adapt over many steps

Example goal: "Research a topic, write a draft, and check sources automatically."

Structure

graph TD
    M[Memory] --> A[Agent]
    T[Tools] --> A
    G[Goals] --> A
    A --> Ac[Actions]
    Ac --> E[Environment]
    E --> O[Observations]
    O --> A

The agent uses its memory, tools and goals to choose an action. The action changes the environment, and the agent observes the result before it chooses the next action.

Agent loop

Agentic AI follows a continuous loop:

graph TD
    P[Plan] --> Ac[Act]
    Ac --> O[Observe]
    O --> R[Reflect]
    R --> RP[Replan]
    RP --> Ac
ComponentRole
PlannerDefines the goal and breaks it into actionable steps
ExecutorCarries out the steps, for example it runs tools, code or API calls
Observer or MemoryRecords the results and the context for the next cycle
ReflectorEvaluates the outcomes, learns from mistakes, and updates the plan

Together, these components make a self-improving feedback loop. This loop is the foundation of autonomy.

Remembering tip

P, A, O, R, then round again: Plan, Act, Observe, Reflect.
Match each step to its part: Planner, Executor, Observer, Reflector.

Opportunities and risks

Opportunities:

  • automation of research, coding, data analysis and workflow design

Strengths:

  • persistence, adaptability, and the ability to handle complex goals

Risks:

  • loss of control
  • unintended actions
  • ethical and safety concerns

From grounding to acting

The 3 topics of this week build on each other:

StageMethodWhat it adds
GroundingRAGFacts from an external corpus
ReasoningCoTStep-by-step logic
ActingAgentic AIActions with tools toward a goal

This trend moves LLMs from tools to collaborators.

Remembering tip

Look, Think, Do: RAG looks up, CoT thinks, the agent does.

Agents that use multimodal LLMs, and their safety risks, are described in Multimodal Agents.

The loop of an agent is explained in terms of who decides each step in Workflow vs Agent.