Define an Agent

An Agent starts as a configuration.

You define what it should do, which Model it should use, and which capabilities it can access.

Minimal Agent shape

name: customer-research
model: deepseek/deepseek-v4-flash
instructions: |
  Research a company and return a concise customer brief.
tools:
  - web_search
  - page_scraping
skills:
  - sandbase/research-report
credentials:
  - SERPER_API_KEY

The exact UI can change, but the idea stays the same:

• Model decides how the Agent thinks.
• Instructions tell it what outcome to produce.
• APIs connect it to real-world data and actions.
• Skills give it reusable know-how.
• Credentials keep API keys and other secrets outside the prompt. SandBase supplies them to the runtime as environment variables.

Test before publishing

Run the Agent from the Console first.

Each test creates a Session. Use its event history to inspect accepted input, Agent responses, errors, and tool activity emitted during the run.

Publish only when it works

When the Agent is tested, publish a selected version as a callable Service, or create a Schedule for manual or recurring runs.

Next steps

• Tools and credentials
• Sessions
• Services
• Schedules