Source

arXiv:2603.22359 — STEM Agent: Self-Adapting, Tool-Enabled, Extensible Architecture for Multi-Protocol AI Agent Systems (March 22, 2026)

Key Contribution

Modular agent framework with MCP-based tool integration and a biologically-inspired skills acquisition system that converts recurring tool-use patterns into reusable capabilities. Multi-protocol adapter design for MCP/ACP/custom protocols.

Relevance to Zeph

Cross-cutting:

• Pattern-to-skill conversion is an alternative implementation path for zeph-skills hot-reload and self-learning (research(skills): ARISE hierarchical RL skill evolution — policy-driven selection outperforms embedding similarity across 7 benchmarks (arXiv:2603.16060) #2398 ARISE overlap)
• Multi-protocol adapter design parallels Zeph's AnyChannel dispatch and MCP/ACP/A2A support
• Biologically-inspired skill promotion (frequency + recency) could augment Zeph's embedding-similarity-based matching

Implementation Sketch

1. Track tool-use patterns in skill_usage_log (tool sequence, context, outcome)
2. When a pattern repeats ≥ N times with positive outcomes, auto-generate a SKILL.md candidate
3. Route candidate through existing skill review pipeline (PR feat(skills): semantic confusability mitigation — category grouping, two-stage matching (#2268) #2402 grouping)
4. Cross-reference with research(skills): ARISE hierarchical RL skill evolution — policy-driven selection outperforms embedding similarity across 7 benchmarks (arXiv:2603.16060) #2398 (ARISE) — may be implementable as a single PR