Back to Projects

Agentic Development Framework (ADF)

A comprehensive environment layer for orchestrating autonomous development agents.

The Challenge

Most “agentic” coding tools are just fancy prompts wrapped in a loop. They lack context, persistence, and distinct roles. I needed a system that treated agents not as tools, but as first-class cognitive participants in a structured development process.

The Architecture

ADF is not just a script; it’s an Operating System for Work. It decomposes development into a two-layer model:

1. The Environment Layer (Ambient)

Six primitives that persist across all projects:

graph TD
    subgraph Environment Layer
        Orchestration --> Capabilities
        Capabilities --> Knowledge
        Knowledge --> Memory
        Memory --> Maintenance
        Maintenance --> Validation
    end
    
    subgraph Project Layer
        Discover --> Design
        Design --> Develop
        Develop --> Deliver
    end
    
    Environment -.-> Project
  • Orchestration: Defines the process (Discover → Design → Develop → Deliver).
  • Capabilities: A registry of tools and skills (MCP servers, plugins).
  • Knowledge: Curated, evergreen learnings (not just raw logs).
  • Memory: Session persistence and cross-project continuity.
  • Maintenance: Self-healing scripts and health checks.
  • Validation: Drift detection and spec compliance.

2. The Project Layer (Ephemeral)

The actual work happens here, moving through strict stage gates: Discover (Intent) → Design (Specs) → Develop (Build) → Deliver (Ship)

The “Builder” Implementation

ADF is built as a set of MCP Servers and Claude Code Plugins.

  • ADF MCP Server: Exposes the environment to agents. An agent can “ask” the environment: “What skills do I have available for Python testing?” or “What did we learn about Next.js hydration errors in the last project?”
  • Codified Governance: Rules are not just suggestions; they are hard-coded constraints (.claude/rules/) that the agent runtime cannot override.

Key Insight

“Context is disposable; Plan is durable.” ADF is designed to wipe the agent’s context window between phases. The agent must re-read the authoritative state (artifacts) to continue. This prevents “context drift” and hallucinations that plague long-running agent sessions.

ready