| name | stem-agent-multi-protocol-architecture |
| title | STEM Agent: Multi-Protocol AI Agent Architecture |
| version | 0.0.3 |
| engine | skillxiv-v0.0.3-claude-opus-4.6 |
| license | MIT |
| url | https://arxiv.org/abs/2603.22359 |
| keywords | ["Agent Architecture","Protocol Unification","Model Context Protocol","Adaptive Skills","Memory Management"] |
| description | Design modular AI agent systems that work across multiple interaction protocols (A2A, AG-UI, A2UI, UCP, AP2) without protocol-specific rewiring. STEM Agent uses biological pluripotency as a metaphor: a generic core differentiates into specialized handlers, tool bindings, and memory subsystems. Validated with 413 tests. Use when building agents that must support diverse interfaces, require adaptive learning from user interactions, or need scalable memory management. |
| category | Research Infrastructure |
Capability Gap
Traditional AI agents lock themselves into a single interaction protocol. Build an agent for REST APIs? It can't easily handle a web UI. Designed for tool calling? It struggles with structured API contracts. This fragmentation forces teams to rewrite agent logic for each new interface, multiplying maintenance burden and creating inconsistency across channels. STEM Agent solves this by decoupling core agent logic from protocol-specific adapters.
Core Abstractions
Undifferentiated Agent Core: A generic reasoning engine that takes goals and produces decisions without assuming any particular I/O protocol. It maintains semantic understanding independent of how requests arrive or responses are delivered.
Protocol Gateway: A unified interface to five interoperability standards (A2A for agent-to-agent, AG-UI for agent-to-GUI, A2UI for agent-to-API, UCP for unified calling protocol, AP2 for agent protocol 2). Each protocol has its own serialization, calling conventions, and error handling, but all funnel through a single gateway that translates between them.
Caller Profiler: Learns continuously from over twenty behavioral dimensions—response latencies, query patterns, error frequencies, preference signals—building a dynamic model of each caller's context and constraints. Enables personalization without explicit user model.
Capability Externalization via MCP: The Model Context Protocol separates domain capabilities (tools, data sources, business logic) from core agent reasoning. New capabilities plug in as MCP servers without modifying the agent core.
Adaptive Skills: A maturation system that crystallizes recurring interaction patterns into reusable skills. When the agent repeatedly performs the same multi-step sequence, it packages it as a skill, making future executions faster and more reliable.
Memory Management: Episodic pruning (forget old sessions), semantic deduplication (recognize repeated concepts), and pattern extraction (identify recurring behaviors) maintain sub-linear memory growth. Critical for long-running agents that can't store all historical context forever.
Design Decisions and Rationale
Decision 1: Protocol Abstraction Layer
Chose a gateway-based architecture that normalizes all protocols into a common internal representation, rather than trying to build a single "universal protocol." Rationale: protocols evolve independently; a normalized internal layer lets the agent evolve separately from protocol evolution. Trade-off: adds a translation layer with small latency overhead, but gains protocol-agnostic reasoning.
Decision 2: Biological Pluripotency Metaphor
Inspired by stem cells differentiating into specialized tissues, the agent core remains generic and unfocused until instantiated for a specific protocol/context. Rationale: enables code reuse (one core logic, many specializations) and simplifies testing (test core logic once, test protocol adapters separately). Trade-off: requires clear boundaries between core and specializations, which increases architectural discipline.
