Drop one hard check from an Artificial Intelligence (AI) agent’s learning loop, and its test score crashes from 17.2 to 6.6. That 62% drop is the main finding in MASkills, a peer-reviewed paper from Arizona State University (ASU), Cisco Research, and the University of North Carolina (UNC). [1] The authors argue the AI field has focused on the wrong fix for too long. [1]

Why Memory Fails Agents

The wrong focus is memory. For a year, giving AI agents more memory has been the standard fix for long tasks. But the MASkills authors explain why this stops working: self-reflection builds memories, but these are hard to trigger, refine, or scale. [1]

A memory just records what happened. It does not tell an agent which rule to use next. [1]

A skill solves this problem. The paper defines a skill as a clear package that tells an agent when to act, how to act, and which tools to use. [1] This makes skills the exact unit an agent can learn.

MASkills grows each agent’s skill library through four simple steps: refining, finding, combining, and cutting. [1][2] This helps a team get sharper across tasks instead of just holding more text in its head. [1][2]

The big picture: The AI agent stack is growing from simply adding context to actively picking skills. Past standards defined how skills are served over the Model Context Protocol (MCP). Now, this paper shows how a multi-agent team learns skills by giving credit to the specific skill that helped. [2]

The surprising part is the safety rail. A skill is a plain text file, so one bad change can silently break an agent’s ability to act. MASkills refuses to save any change that fails a strict test on fresh data, rolling back the system if it fails. [2]

By the numbers:

  • 76.3 F1 score: MASkills hit this accuracy mark on the HotpotQA reasoning test. [2] It beat baselines like MultiPersona (69.2) and Chain-of-Thought with Self-Consistency (68.9). [2]
  • 17.2 to 6.6: The drop in the LoCoMo reasoning test score when researchers cut the rollback gate. [2]
  • 35.3: The system’s score on the General AI Assistant (GAIA) test, compared to 28.2 for standard agents. [2]
  • 3 setups: The system works across peer, central, and tiered agent teams. [2] Peer teams win on reasoning, while central teams win on long-term memory. [2]

The Skill Playbook

MASkills offers a clear guide for builders making multi-agent systems. Three core moves stand out.

The playbook:

  • Optimize skills, not context: Replace huge memory stores with clear skills that carry exact rules for when to use them. Load the summary first, then load the full skill file on demand to keep the prompt short. [1][2]
  • Assign credit by skill: Use a central Large Language Model (LLM) judge to grade how much each skill helped the team. This ensures updates fix the exact skill that worked. [2] If a failure maps to no known skill, that means it is time to build a new one. [2]
  • Gate edits with rollbacks: Save a skill change only if it passes a fresh test. [2] If it fails, restore the old version to stop one bad guess from breaking a working agent. [2]

Limits of Skill Libraries

The catch: MASkills is a research tool, not a ready-to-use product. The tests rely on helpful agents with fixed roles. [2] The authors list hostile and open-world setups as clear limits. [2]

Skill libraries also face scaling limits. As libraries grow, finding and combining skills gets harder, which remains an unsolved problem. [2] Finally, fixing text with an LLM judge is still based on chance. The hard check catches bad changes, but it cannot make the judge perfectly consistent.

Go deeper:

  • The Big Shift: A new paper called MASkills shows multi-agent systems improve faster when builders focus on clear skills instead of piling up memory.
  • Why It Matters: The paper proves that strict safety checks are vital for AI agents, as cutting the rollback check drops reasoning performance by 62%.
  • The Winning Moves:
  • Optimize skills, not context: Give agents clear rules loaded on demand to replace bloated memory stores.
  • Assign credit per skill: Use a central judge to grade each skill’s impact so updates target the right behavior.
  • Gate edits with rollback: Save a skill change only after it passes a test, reverting it if it fails.
  • The Catch: The tests assume agents cooperate in fixed roles, leaving open-world spaces untested and skill library scaling limits unsolved.

Related reading

Sources

[1] https://arxiv.org/abs/2609.02094 — MASkills: Continual Skills Optimization for Multi-Agent LLM Systems (abstract; EMNLP 2026 Findings, submitted 2026-09-02) [2] https://arxiv.org/html/2609.02094v1 — full paper (Tables 1–3, ablations, topology results) [3] https://github.com/DaRL-GenAI/MASkills — code