A Practical Introduction to Agentic AI

A Practical Introduction to Agentic AI

• Sudip Shrestha, Lead AI Engineer at ASI Government
  • linkedin
  • website
  • slides
  • ASI Government

NoteNotes

• Main objective: Build a working agentic artificial intelligence application from scratch in Google Colab, ending with a simple Gradio demo app.
• Core distinction: chatbot vs. agent
  • A chatbot mainly responds to prompts.
  • An agent can decide, act, call tools, loop through steps, and move a workflow forward without constant human prompting.
  • The underlying large language model may be the same; the difference is the workflow, tools, state, and control logic wrapped around it.
• Why agentic AI is becoming practical now
  • Tool calling has become more reliable.
  • Frameworks such as LangGraph, CrewAI, AutoGen, and OpenAI’s SDK have matured.
  • There is growing demand for applications that do more than produce text: they execute tasks, revise outputs, and integrate with systems.
• Key agentic AI capabilities discussed
  • Tool use through Python functions or APIs.
  • Task decomposition into steps.
  • Self-correction through critique and revision.
  • Conditional routing based on intermediate results.
  • Human-in-the-loop review when autonomy is risky.
  • Observability and tracing for debugging.
• Important limitations and risks
  • Agents can become expensive or slow if they loop too much.
  • Loops need maximum iteration caps.
  • Results are nondeterministic, so evaluation must be systematic rather than based on one successful run.
  • Tool descriptions and docstrings matter because the model uses them to decide which tool to call.
  • Security requires least-privilege access, authentication, and careful API control.
  • Debugging is harder because the model, tools, prompts, and graph logic all interact.
• LangGraph concepts introduced
  • State: shared memory passed through the graph.
  • Nodes: Python functions that read and update the state.
  • Edges: fixed transitions between nodes.
  • Conditional edges: decision functions that choose the next step at runtime.
  • Loop caps: safeguards that prevent infinite or costly cycles.
• Most important technical idea
  • Conditional edges turn a fixed pipeline into an agentic workflow.
  • Example: a should_revise function checks a score; if the score is high enough, the workflow stops, otherwise it routes back to revision.
• Hands-on build structure
  • Part 1: Start with a plain large language model call, then add tool calling.
  • Part 2: Build a LangGraph stateful self-correcting agent.
  • Part 3: Wrap the system in a Gradio interface and store results in SQLite.
• Tool-calling example
  • Shrestha demonstrates simple Python tools such as add and multiply.
  • These functions are wrapped so the model can decide when to call them.
  • The docstring acts like prompt engineering inside code: it tells the model when and how to use the tool.
• Self-correcting LinkedIn post agent
  • The agent takes a topic and drafts a LinkedIn post.
  • An evaluator scores the post out of 10.
  • If the score is below the threshold, the agent revises the draft.
  • If the score passes the threshold, the workflow ends.
  • The process stores the topic, score, revision count, and feedback.
• Structured output
  • Pydantic is used to force model outputs into a predictable schema.
  • This makes evaluation and routing more reliable because the graph can depend on fields such as score and feedback.
• Application layer
  • Gradio is used for a lightweight demo interface.
  • SQLite stores generated posts and metadata.
  • Shrestha notes that Gradio and Streamlit are useful for demos, while production apps may need frameworks such as Next.js.
• Production considerations
  • Choose models according to task complexity, cost, and modality.
  • Use stronger models only when the task justifies the cost.
  • Keep evaluation criteria specific; vague rubrics lead to unreliable outputs.
  • Multiple specialized agents can collaborate, but each should have a narrow role.
• Final takeaway
  • Agentic AI is not just “a better chatbot.”
  • It is a workflow architecture where a language model uses tools, maintains state, evaluates its own work, and conditionally decides what to do next.