Add an optimization agent

Implement a selection policy when grid search does not fit your search space or compute budget.

Use the public lifecycle

Subclass BaseAgent from cloudai.configurator.base_agent and provide its configuration model. Start from GridSearchAgent for a complete working implementation, then change selection and feedback handling.

• get_config_class declares the configuration model.

• configure prepares the policy’s search state.

• select_action(observation=...) returns a step number and action mapping, or None to stop.

• update_policy(feedback) receives the action, reward, current and previous observations, trial index, and done flag.

The base loop uses max_steps as its iteration bound. Implement budget accounting and early termination in the policy when required. Grid search sets this bound to the number of combinations and ignores observations. Agents that override run must preserve its success/failure contract.

Package the integration

CloudAI discovers installed packages through the cloudai.agents entry-point group. For example, in your package’s pyproject.toml:

[project.entry-points."cloudai.agents"]
my_search = "my_package.agent:MyAgent"

Install that package into the environment running CloudAI and select agent = "my_search" in the case configuration. Use agent_config for your model’s parameters. See Agent and search settings for shared settings.

Verify the policy

Test action validity, observation ordering, deterministic seeding where supported, missing metrics, and termination. Demonstrate that the stopping policy respects the budget you document. Distinguish a recoverable workload failure from an unexpected agent bug. Inspect the saved trajectory and selected configuration, not just the returned exit code.