Configure inputs on an eval
Open or create an eval in Agent Experience. Choose the environment and resources the agent should use:
In the eval editor, the repository and persona appear under Environment. Skills, MCP servers, and credentials appear under Resources. These selections are saved with the eval, so you can run it again with the same configuration. You do not need a
gauge.json file to configure an eval in the platform.
Use setup and teardown scripts when the task needs preparation or cleanup, such as creating test data before a session. See Configure runs for configuration and Credentials for service access.
Choose inputs that match your users
Give the agent what a developer would have when using your product. For an integration task, that may be an existing application, your setup Skill, and access to a test account. For a task that measures product discovery, supplying a product-specific Skill can change the question: the agent already has instructions for that product. Keep the task and criteria separate from the inputs. A Skill can explain how to use your API; a criterion checks whether the finished integration works. Selecting a Skill or MCP server makes it available to the agent, but does not guarantee the agent will use it successfully. Agents can also read documentation and install publicly available packages while completing a task. You do not need to declare every page or package they might discover as a project input.Compare changes to an input
To test whether an input improves the result, keep the task, criteria, models, and other inputs consistent. For example, compare sessions with the current integration Skill against sessions with a revised version. Inspect both the outcome and the recorded behavior to see what changed. Pin a repository commit when you need a repeatable starting point. Check the selected resource versions and credentials before rerunning an eval. See Improve agent experience for the improvement workflow.Supply changes through GitHub
The GitHub integration usesgauge.json to describe product inputs supplied by a repository or its build and preview workflows. This is how a PR check can test an unpublished package, a revised Skill, supporting files, a native CLI, or a docs preview.
Those declarations describe how the integration delivers the candidate changes to the evaluation. The GitHub PR checks guide covers the supported types, sources, and case selection. See the CLI guide for supplying local build outputs.