atomic-agent-designer

Designing atomic agents through structured dialogue uses a five-phase workflow to produce an actionable Agent Card. It guides you from idea to specification using MAP, GROUND, CONTRACT, VALIDATE, and AGENT CARD phases.

The contract-first approach defines inputs and outputs before describing the agent's process, establishing explicit boundaries. This ensures predictable behavior by enforcing constraints and one DIKW transition per agent upfront.

Mapping DIKW transformations and Kolb learning cycles is supported through task-mode and process-mode workflows. These dual flows integrate real-world prompts to align learning cycles with data, information, knowledge, and wisdom transitions.

The atomic agent design methodology enforces three-example validation to ensure reliability. Validating examples before generating an Agent Card guarantees the agent handles expected inputs correctly within its defined boundaries.

A living Agent Card is an evolutionary artifact that updates after real-world use via a Kolb journal. It tracks experiential learning to continuously refine the agent's specification based on actual performance and feedback.

Composing multiple atomic agents is supported because each agent handles exactly one DIKW transition. This composable structure allows you to chain agents together to manage complex workflows while maintaining strict individual boundaries.