Skip to main content
The best first version of an agent is narrow enough to test and useful enough that someone will read its output.

Design the workflow first

Write down:
  • Outcome: What decision, report, or updated state should exist?
  • Owner: Who reviews quality and responds to failures?
  • Target: Organization-wide, selected accounts, or a dynamic group?
  • Inputs: Which records, files, systems, or prior memory are required?
  • Output: Narrative, structured fields, a table, or another defined result?
  • Cadence: On demand or scheduled in the organization’s timezone?
  • Boundary: Which writes, if any, are allowed? When must the agent stop?
  • Success: How will a reviewer know the run was useful and correct?

Create the draft

1

Open Agent Library

Choose New Agent or the available create action.
2

Name the job

Use a verb and outcome, such as “Review backup exceptions” rather than a department name.
3

Write the instructions

State the workflow, evidence requirements, output format, stop conditions, and what not to infer.
4

Choose the model

Use Fast for routine, well-defined extraction and reporting. Use Smart for workflows with more ambiguity or cross-source reasoning.
5

Keep it inactive

Save the first version as inactive while you configure tools and test.

A durable instruction structure

  1. Purpose and audience
  2. Inputs and source precedence
  3. Ordered workflow
  4. Tool-use boundaries
  5. Required output
  6. Evidence and confidence
  7. Stop and escalation conditions
If two outputs have different owners or failure behavior, they are often better as two agents. Smaller agents are easier to test, permission, and maintain.
Next, configure the agent.