Understand the agent lifecycle and authoring surface
Creating an agent is an iterative process. You define its purpose, configure its behavior, test it, publish it, and improve it based on results.
Follow the agent lifecycle
Follow these five stages:
- Create the agent and establish its purpose.
- Build its instructions, context, capabilities, model, and boundaries.
- Test behavior in Preview and with repeatable evaluations.
- Publish the agent to selected channels.
- Monitor activity and use evidence to improve the agent.
Evaluation and monitoring can reveal an unclear instruction, overlapping tools, an unsupported answer, or a missing boundary. Return to the agent, make a focused change, test it again, and publish the update.
Use the authoring tabs
The agent workspace contains four main tabs that support this lifecycle:
- Build: Configure the agent's instructions, knowledge, tools, skills, model, connected agents, and memory.
- Preview: Try representative requests while you build the agent.
- Evaluate: Run a defined set of test cases to measure quality and compare changes.
- Monitor: Review recent tasks, accessed files, and activity after the agent is in use.
A Preview conversation shows one example of how the agent responds. An evaluation runs a defined set of cases so you can measure quality more consistently. After publishing, use Monitor to review actual activity and identify improvements.