Chat, workflow, or agent: choose how much work to delegate

Understand the difference between a conversation, a repeatable sequence, and an agent that can choose actions.

The key question is what the tool can do after it answers.

Look at the available actions

A text conversation can propose a plan. A tool-enabled assistant might also read files, run calculations, create an artifact, or interact with another service. Product labels vary, so inspect the actions and permissions in the application you are actually using. “Agent” is not a guarantee of any specific capability. Decide what result you need and which actions are necessary to produce it.

Use a sequence when the steps are known

Consider turning meeting notes into a draft action list. The sequence may be stable: extract decisions, identify owners, flag missing dates, and format the draft. A predefined workflow gives you clear places to inspect the result. If the work instead involves exploring an unfamiliar codebase and choosing which files to change, an agent may need to decide its own intermediate steps.

Make the first delegation small and reviewable

Choose a bounded output: one report, one folder inventory, or one feature prototype. Give the relevant sources and define what counts as complete. For a folder inventory, completion might mean every visible file is listed with its size and a proposed category, with nothing moved. Ask for a sample if a wrong interpretation would cause substantial rework.

Evaluate the complete result

An assistant saying “done” is a status report, not proof. Open the artifact, inspect the relevant behavior, and compare it with your acceptance criteria. If the tool made changes, review what changed. Record failures as specific cases to improve the workflow. Start with the simplest arrangement that handles the task reliably; adding more moving parts also adds more places to inspect.

A worked example

Original teaching example, not a recorded model run.

Vague starting point: Be my autonomous assistant and organize everything.

A more useful brief

Inventory the files in this project folder and propose a grouping. Do not move, rename, or delete files. Show the inventory and five example classifications first. Completion means every file is accounted for or explicitly marked unreadable.

Why this helps: This establishes the artifact, scope, action boundary, and a way to verify completion.

Your turn

Use the workflow planner to describe one recurring task. Decide whether you need a conversation, a fixed sequence, or an agent.

You are looking for this: You can name the actions needed and the evidence that would show the task is complete.

A quick check

A task always follows the same four steps. What is a sensible starting point?

1. A simple repeatable workflow with checks between steps
2. As many agents as the tool allows
3. No instructions because autonomy is always better

Answer: A simple repeatable workflow with checks between steps

Start with the simplest approach that handles the task and its exceptions.

Sources and how to read this guide

Examples and exercises are original teaching material, not recorded model runs, measured time savings, or certifications. Official sources were consulted September 6, 2026. Product behavior can change.

• Anthropic · Building effective agents: Distinguishes predefined workflows from systems that choose their own steps.

Take the idea further

• Give an AI agent a clear brief and sensible permissions
• Test an AI workflow before you depend on it

All learning paths | Practice challenges.

Related Power of AI pages

Keep reading with Start here, everyday uses, the tool guide, the writing workshop, and sources and standards.