MAS-SBR Mesh

Architectural Standard: MAS-SBR Mesh

1. Abstract

The MAS-SBR framework defines a system where intelligence is not a property of a single agent, but an emergent property of a Skill-Based Route. By decoupling “Skills” (atomic functions) from “Agents” (logical sequencers), the architecture allows for a many-to-many relationship that optimizes for task-specific logic patterns.

2. Structural Components

2.1 The Skill Atom ($S$)

The smallest executable unit of work. Skills are stateless and idempotent where possible.

• Examples: Data_Fetch, Syntax_Check, Tone_Shift, Cross_Reference.

2.2 The Agent Node ($A$)

An Agent is a “sequencer” of specific Skill Atoms. Its unique value is its Execution Vector ($V$).

• Agent A Vector: $S_1\to S_2\to S_3$ (Deductive Logic)
• Agent B Vector: $S_3\to S_1\to S_2$ (Inductive Logic)

2.3 The SBR Mesh (The “Web”)

The topology is a Directed Acyclic Hypergraph.

• Nodes: Agents and Skill Repositories.
• Edges: Data handoffs defined by the routing logic.
• M:M Relation: Multiple tasks can trigger the same agent; a single agent can call upon multiple skill-sharing peers to fulfill a missing atom in its vector.

---

3. Dynamic Routing Logic

3.1 The Dispatcher (The Router)

The Dispatcher does not look for an “Agent Name”; it looks for a Logic Profile.

1. Requirement Analysis: Parses the user prompt for required Skill Atoms.
2. Permutation Matching: Identifies which Agent Vector ($V$) produces the most appropriate “State Transformation” for the goal.
3. Cost-Benefit Selection: Evaluates agent availability (latency) vs. historical success rates for that specific sequence.

3.2 State Handover Protocol (SHP)

To maintain the many-to-many net, data must be wrapped in a Context Carrier as it moves through the mesh:

• Payload: The actual data/result.
• Trace: The list of skills already executed ($S_{history}$).
• Intent: The remaining skills required to satisfy the route ($S_{remaining}$).

---

4. Operational Comparison

Feature	Legacy Hierarchical MAS	MAS-SBR Mesh
Connectivity	1:Many (Tree)	Many:Many (Web)
Redundancy	Low (Single point of failure)	High (Skill overlap across nodes)
Logic Type	Fixed/Declarative	Permutational/Procedural
Scalability	Linear (More tasks = More agents)	Exponential (Skill reuse allows new logic patterns)

---

5. Maintenance & Evolution

5.1 Adding a New Skill

When a new Skill Atom is registered to the central library, all agents in the mesh immediately gain the potential to include it in their vectors. No hard-coding of connections is required; the SBR Dispatcher simply acknowledges the new capability in its next routing cycle.

5.2 Conflict Resolution

If two agents, Agent X $(S_a,S_b)$ and Agent Y $(S_a,S_b)$, both claim the same task, the mesh uses Telemetry-Based Selection:

• Success Delta: Which agent’s specific LLM/Code implementation of $S_a$ produces higher quality output for this specific domain?
• Load Balancing: Route to the agent with the lowest current token-queue depth.

---

6. Visualization of the Many-to-Many Net

graph LR
    Task1[Code Review] --> Router{SBR Dispatcher}
    Task2[Feature Draft] --> Router
    Router --> AgentA((Agent A: S1, S2, S3))
    Router --> AgentB((Agent B: S2, S3, S1))
    AgentA -.-> Skill_S2{Skill S2 Repository}
    AgentB -.-> Skill_S2

---