By KOH NOMAD · AI & digital business
Compare three working patterns
For a fixed workflow, you decide the sequence: receive a form, check required fields and create a review item. The path is predictable and can often be expressed as explicit rules.
An assistant may draft a response when a person asks. An agent can decide which permitted tool or intermediate step to use as the task develops. These are practical distinctions, not universal marketing definitions. Products described as agents may contain a mixture of these patterns.
Use the least complexity that solves the task
Anthropic's engineering guidance distinguishes predefined workflows from systems in which a model directs its process and tool use. It recommends starting with simple approaches and adding complexity when the task warrants it.
If every enquiry follows the same routing rule, a workflow may be easier to inspect. If research requires following different leads and deciding which information is missing, a bounded agent may be useful. Test the need for flexibility rather than using the word agent as a quality label.
Define a job the system can finish
An illustrative research agent could inspect approved public sources, identify unanswered questions and prepare a briefing for a human. Its output would be a dated note with evidence, not permission to contact anyone or buy a service.
Write down the objective, permitted sources and tools, required evidence, completion condition, budget and stop condition. Decide who takes over when information conflicts or the system cannot complete the task.
Keep authority separate from intelligence
A model's ability to suggest an action does not establish permission to perform it. Define which steps can run automatically, which need review and which are unavailable. Apply the same access boundaries to every tool the system can use.
For example, allow a draft summary but require approval before an external message, a published claim or a commercial commitment. Give access only to the records needed for that job and keep credentials out of prompts and visible logs.
Test the evidence and the failure paths
Use representative tasks and compare results with a manually checked answer. Test missing data, conflicting sources, duplicate triggers and tool failures. Record whether the result was accurate, useful, within budget and complete.
A queue item is not proof that a task ran. A draft is not a sent message. A successful tool request may still need confirmation from the destination. Make those states visible so a person can intervene without guessing.
Begin with a controlled pilot
Choose one narrow task and run it on examples before connecting it to important operations. Review errors and the real cost, then decide whether more autonomy is justified.
Use the guides below to turn the idea into a workflow brief. The project CTA can help you describe the stage you are at; it does not activate an agent or an automatic trading system.
Official sources
Sources checked: 2026-11-16. Recheck current rules before acting.
Your quick checklist
- Does the task need a fixed process or model-directed decisions?
- What objective and completion condition can be checked?
- Which tools and data are permitted?
- Where are human approvals required?
- What are the cost and stop limits?
- How are errors, retries and evidence reviewed?
Not sure where to start? Find your next business step.
Find a place for your next workday
Explore public city guides, then confirm the current conditions with the venue.
