Composable Pipeline Architecture

The Composable Pipeline Architecture is how Maudel assembles instructions for every pipeline stage. Instead of hardcoded prompts, each stage is defined by markdown files that are composed into a layered system prompt at runtime.

Modular Stage Design

Each pipeline stage has a modular design composed of:

• Expert agents with domain-specific learnings and tools
• Skills honed for stage-specific tasks and artifact creation
• Quality gates with pluggable evaluators for real-time supervision
• Deterministic progression rules defined by the user and observed by the orchestrator
• 9-layer Action prompt specifically constructed to yield accurate, efficient results

9-Layer System Prompt Assembly

At runtime, Maudel assembles a system prompt from nine layers organized into three cognitive zones:

Primacy Zone (Strongest Recall)

1. Agent Identity & Scope — who the agent is and what it can do
2. Pipeline Instructions — global pipeline rules and constraints

Reference Zone

3. Stage Instructions — specific directives for this stage
4. Reference Skills — available capabilities and domain knowledge
5. Mode Instructions — collaborative vs. autonomous behavior
6. Special Instructions — project-specific overrides

Recency Zone (Second Strongest Recall)

7. Story-Specific Instructions — context for the current task
8. Expected Outputs — what the agent must produce
9. Upstream Context — relevant outputs from prior stages

This architecture leverages cognitive science principles — critical instructions land in primacy and recency positions where LLM attention is strongest.

Pipeline Templates

There are 6 system pipelines available out of the box, but users can create custom pipelines by selecting any combination of stages and configuring the agents, skills, tools, quality gates, and progression rules for each stage.

This allows for flexible pipeline configurations tailored to different project needs while maintaining deterministic execution and governance.

Input Hydration

Each stage declares its required inputs. At runtime, the orchestrator matches declared inputs to actual runtime values and injects them into the prompt. This ensures agents always have the context they need — and never more than they should to avoid context bloat and context rot.

Extensibility

Soon adding a new pipeline stage will be as simple as:

Upload a DSL file defining the stage’s agents, skills, tools, quality gates, and progression rules and zip, boom, bam, it’s live. No code changes. No redeployment. The composable architecture makes the pipeline a living, evolving system. Stay tuned…