Skip to content

Evolutionary Strategy

AgentCommander manages the lifecycle of experiments using a structured evolution logic.

Default Strategy: B/L/S (Branch/Level/Step)

The default workflow (default_graph.json) implements a heuristic evolutionary tree:

  • Branch (B): Represents major conceptual directions or distinct lineages.
  • Level (L): Represents depth of optimization. When an experiment succeeds (improves the metric), the system spawns a new generation at Level + 1, inheriting the successful traits.
  • Step (S): Represents horizontal trial-and-error. When an experiment fails, the system retries at Step + 1 with a modified hypothesis, keeping the Level constant.

Customization & Parallelism

It is important to note that B/L/S is simply the default template.

Flexible Graph Architecture

The entire logic is defined in the visual workflow editor. You are free to redesign this graph to implement: * Genetic Algorithms: Selection, Crossover, Mutation nodes. * Bayesian Optimization: Logic nodes that update hyperparameters based on past results.

Parallel Exploration

The underlying engine supports Parallel Execution. You can design workflows where multiple agents explore different branches simultaneously, or where a "Manager Agent" spawns multiple "Worker Agents" to solve sub-problems in parallel. This scalability allows AgentCommander to leverage massive compute resources for accelerated discovery.