Workflow vs agent
Who decides the next step: your code or the model.
What you’ll learn
Four ideas, one module.
The agent loop
Observe, decide, act, update state. Repeat until done or stopped.
Errors compound
Why a chain of mostly good steps can still fail as a whole.
Autonomy by design
Where a human approves, and when a plain workflow is the better choice.
Applied exercise
Redesign four use cases
Redesign four use cases as workflows, partially agentic systems and autonomous agents, then compare risk.
- Customer refund requests
- Weekly sales report
- Competitor research brief
- Night-time incident response
Cheat sheet · for your projects
Autonomy is a cost: give the model a decision only where code can’t cover the cases.
Ask on every project
- Who decides the next step here: our code, or the model?
- Are the steps the same for every input, or do they depend on what it finds?
- What stops the loop: a step limit, a budget, a timeout or a human?
- If step 3 is wrong, which later step or person notices it, and when?
- If a run fails halfway, what state do we need to resume it or undo it?
Rules of thumb
- Make known steps fixed code; keep model decisions for the parts that truly vary.
- Set step, budget and time limits in code, and decide what happens when one is hit.
- Check key results before they enter state; later steps will treat them as facts.
- Put approval on few, high-risk actions, and show the approver the evidence.
Decision guide
How much autonomy?
- If the steps are the same for every input→ fixed workflow; the model only fills in steps
- If inputs fall into a few known paths→ router: one model decision, then fixed code
- If the path depends on what it finds and tools are read-only→ bounded agent with caps
- If an action sends, spends, deletes or changes production→ a human approves it first
- If an action can’t be undone and approval isn’t practical→ don’t expose it as a tool
- If you can’t yet measure whole runs→ stay one level less autonomous until you can
Leave at the door
Five myths this module breaks.
| Myth | What’s actually true |
|---|---|
| “More agentic is always better.” | More autonomy adds variance, cost and risk too. |
| “The agent takes the action.” | The model requests it. Your system executes it. |
| “It made a plan, so it stays on track.” | A plan is text. Only checks keep it on track. |
| “More agents give better answers.” | Each hand-off adds cost and a chance to fail. |
| “A human in the loop makes it safe.” | Only if they see the evidence and can say no. |
Next · Module 08