When to use an agent
Use an agent when you need a repeatable outcome across several steps, sources, or applications. Agents are most useful when they must interpret inputs, choose a path, and produce an answer or action.
An agent is a good fit when you need to
- Combine information from multiple documents, conversations, or connected systems.
- Apply judgment or conditional logic to changing inputs.
- Run a sequence of read and write actions.
- Start work from a schedule, message, or supported system event.
- Produce a consistent output such as a brief, triage summary, recommendation, or follow-up plan.
- Share a repeatable process with a team and improve it over time.
Start with a narrow outcome. Add tools and write actions only when they are necessary for that outcome.
Choose a simpler experience when
| If the task is... | Consider using... | Why |
|---|---|---|
| Finding an answer in company content | Search or Assistant | You need retrieval and explanation, not a reusable execution plan. |
| One well-defined action in a connected app | A tool | A single action is easier to understand, permission, and maintain. |
| A fixed process with no need for AI judgment | A deterministic workflow or Workflow mode | Explicit steps provide more predictable execution. |
| Task-specific guidance that should be reused in Assistant and agents | A Skill | Skills package instructions and expertise without owning a run or lifecycle. |
| A high-risk decision that requires human accountability | A reviewed process with human approval | Keep the decision with the responsible person and use automation only for preparation. |
Examples by department
- Sales: Research an account, compare recent activity with the account plan, and draft a meeting brief.
- Support: Classify an incoming request, find the relevant troubleshooting content, and prepare a response for review.
- Engineering: Gather the failing test output, find related incidents, and suggest likely owners and next steps.
- Human resources: Compare an onboarding request with the approved checklist and identify missing steps without making employment decisions.
- Finance: Collect inputs for a monthly close summary and flag missing data for an analyst to review.
Examples by industry
- Software: Triage an incident by combining alerts, runbooks, ownership data, and recent tickets.
- Financial services: Prepare a research packet from approved internal sources while leaving regulated decisions to a qualified reviewer.
- Healthcare: Organize non-clinical service requests and route them to the correct team without making clinical recommendations.
- Retail: Summarize store operations issues and prepare tasks for regional managers.
- Manufacturing: Compare a production exception with maintenance procedures and draft an escalation summary.
Examples by task pattern
Agents work well for research and synthesis, triage and routing, drafting, monitoring, and multi-step updates. For example:
- Read a support ticket and related product documentation.
- Determine which issue pattern matches the ticket.
- Draft the recommended next steps.
- Ask a person to approve any external update.
If the task is only step 1 or only step 3, use search, a Skill, or a tool instead of an agent. If every input follows the same rules, use Workflow mode or another deterministic workflow.
Set boundaries before you build
Define the agent's owner, inputs, approved knowledge sources, allowed tools, write actions, and escalation path. Do not give an agent broad access just because a task might need it. Test representative and edge-case inputs before you publish the agent.