Module 1 — Introduction to Agentic Workflows

Lesson notes

What agentic AI means

A direct LLM call maps one input to one output. An agentic workflow breaks an outcome into steps and allows the model to create, inspect, revise, call tools, or delegate before returning a result. The useful unit is the whole workflow, not an isolated prompt.

Degrees of autonomy

Autonomy is a spectrum:

Deterministic pipeline — code fixes the order and the model fills narrow steps.
Router — the model selects one of approved paths or tools.
Adaptive workflow — the model may repeat or reorder bounded steps.
Planner — the model proposes and executes a task-specific plan.
Highly autonomous system — it pursues a broad goal over many steps.

Choose the lowest level that can solve the problem. More autonomy increases coverage, but also cost, latency, nondeterminism, and the number of failure modes.

Why workflows help

Decomposition makes complex work easier to reason about.
Intermediate results expose errors before they contaminate the final answer.
Tool results ground the model in current or private facts.
Feedback and revision raise quality on tasks that can be checked.
Specialized stages can use different prompts, models, budgets, and permissions.

Good applications

Agentic workflows are valuable when work is multi-step, the next action depends on observations, tools contain necessary information, and quality can be evaluated. Examples include research, coding, document processing, customer support, analytics, and operations. A fixed function is better when rules are stable and exhaustive.

Task decomposition

Start from the deliverable and identify observable stages:

Goal → required evidence → subtasks → tools/data → checks → final synthesis

Each step should have an input, output contract, owner, permitted actions, and failure behavior. Keep deterministic work—validation, arithmetic, formatting, authorization—in normal code where possible.

Evaluations

Define success before increasing complexity. Build a small dataset with normal cases, edge cases, and adversarial cases. Evaluate the final result and important intermediate components. Useful dimensions include task completion, factuality, instruction following, tool correctness, safety, latency, and cost.

Design-pattern overview

Reflection: generate, critique, revise.
Tool use: observe or act through controlled interfaces.
Planning: dynamically choose and order steps.
Multi-agent: distribute roles and synthesize their work.

Research-agent application

flowchart LR Q[Question] --> P[Research plan] P --> S[Search sources] S --> X[Extract evidence] X --> G[Identify gaps] G -- more evidence needed --> S G -- sufficient --> W[Write cited answer] W --> V[Verify claims and citations]

Minimum safeguards: source allowlist where appropriate, citation-to-claim checks, a search budget, explicit uncertainty, and no invented evidence.

Check yourself

Can a deterministic pipeline solve this task more safely?
What observations may change the next action?
What is the smallest useful autonomy level?
How will I know the workflow improved over a direct prompt?

Module 1 — Introduction to Agentic Workflows