ADR-009: AI Agent Configuration via Prompt Management DSL

Status: Accepted Version: 1.0 Date: 2025-10-01 Supersedes: N/A Related ADRs: N/A Related PRDs: PRD-012, PRD-013

Context

[!NOTE] Evolution Path: This ADR is foundational but is extended by ADR-026: Cognitive Supply Chain (SOPs). While ADR-009 defines the DSL, ADR-026 defines the versioned supply chain for those prompts.

To provide a flexible and modular way to configure AI agent behaviors, prompt templates, and skill composition.

Decision

Utilize a Prompt Management DSL for configuring AI agents, supporting modular prompt templates, context injection, and skill composition (analogous to LoRA/Adapter patterns).

Rationale

This DSL allows for precise control over AI agent behavior, enabling dynamic adaptation to context and task requirements. It facilitates the integration of various context sources (knowledge graph, domain services, user conversation) and the definition of tool-use capabilities, making agents more effective and manageable.

Alternatives Considered

Hardcoded Prompts

Rejected - Lack of flexibility, scalability, and difficulty in managing complex agent behaviors.

Manual Prompt Engineering

Rejected - Labor-intensive, error-prone, and does not support dynamic adaptation.

Constraints

• MUST use Prompt Management DSL for AI agent configuration
• MUST support modular prompt templates
• MUST support context injection from knowledge graph, domain services, and user conversation
• MUST support skill composition (analogous to LoRA/Adapter patterns)
• MUST NOT use hardcoded prompts

Quality Attributes

• Highly configurable and adaptable AI agents
• Improved prompt engineering efficiency
• Better control over agent behavior
• Support for continuous learning
• Dynamic adaptation to context and task requirements

Bounded Contexts Impacted

• Cognitive Extension Layer
• AI Agent Runtime
• Knowledge Layer

Consequences

Positive

• Highly configurable and adaptable AI agents
• Improved prompt engineering efficiency
• Better control over agent behavior
• Support for continuous learning

Negative

• Requires development of the Prompt Management DSL and associated tooling
• Potential complexity in managing numerous configurations

Additional Notes

✅ MVP