TL;DR

The secret to reliable AI agents is not more parameters, but Externalization. This paper argues that modern agent design is shifting from "Model-Centric" to "Harness-Centric," moving cognitive burdens—like memory, procedural expertise, and interaction rules—out of the model’s weights and into a managed infrastructure called the Harness.

Background: The Outward Migration of Intelligence

For years, the industry assumed that "smarter weights = better agents." However, even the most powerful models fail at long-term consistency. The researchers track a historical arc of capability:

1. Weights (2022-2023): Knowledge is frozen in parameters. Hard to update, easy to hallucinate.
2. Context (2023-2024): Using RAG and Prompting (Chain-of-Thought) to "stage" cognition.
3. Harness (2025-2026): Intelligence is distributed across persistent memory, skill registries, and standardized protocols.

Figure 1: Just as humans moved from internal thought to writing and digital computation, LLM agents are moving from internal weights to external harnesses.

The Core Architecture: Memory, Skills, and Protocols

The paper breaks down "Agency" into three externalized modules that transform the model's task:

1. Externalized State: Memory

Instead of cramming everything into a fragile context window, the harness uses a Hierarchical Memory Architecture.

• Transformation: From Recall (hard) to Recognition/Retrieval (reliable).
• The Design: Modern systems use "OS-style" memory management, swapping "hot" context (active work) for "cold" storage (historical episodes).

2. Externalized Expertise: Skills

A model shouldn't have to "invent" a workflow every time.

• Transformation: From Improvised Generation to Structured Composition.
• The Artifact: "Skill files" (like SKILL.md) encapsulate procedures, decision heuristics, and safety constraints. The agent simply loads the expertise required for the task.

3. Externalized Interaction: Protocols

Interaction with tools (APIs) or other agents is often the point of failure.

• Transformation: From Ad-hoc Language to Structured Contracts.
• The Solution: Standards like the Model Context Protocol (MCP) or Agent-to-Agent (A2A) protocols ensure that communication is typed, validated, and secure.

Methodology: The Harness as a Cognitive Environment

The "Harness" is the runtime that hosts these modules. It isn't just "glue code"; it's a Cognitive Environment that implements:

• Agent Loops: The "Perceive-Plan-Act" cycle.
• Sandboxing: Forcing the model to work in a safe, isolated area.
• Observability: Creating a "black box recorder" of every decision.

Figure 2: The community focus is shifting from the 'Agent Core' (LLM) to the 'Harness' (Infrastructure).

Deep Insight: The Cerebrum vs. The Cerebellum

In a fascinating extension to Embodied AI (Robotics), the authors suggest we are seeing a "Cerebrum–Cerebellum split."

• The Cerebrum (LLM Agent): Handles high-level reasoning and task decomposition.
• The Cerebellum (VLA Models): Handles fast, reactive motor control (e.g., grasping an object). By externalizing motor control as a "Skill," the high-level brain is freed to focus on the goal, making robots significantly more robust.

Critical Analysis & Conclusion

This paper changes the "SOTA" definition. A system’s power is no longer just its benchmark score on a static test, but its externalization quality.

The Trade-off: Externalization adds latency and context overhead. If you externalize too much, the model spends all its time reading manuals rather than working. The future of AI research will be "Partitioning": deciding exactly which 10% of a task should stay in the model's brain and which 90% should be moved to the harness.

Final Takeaway: We are entering the era of "Distributed Agency." If you want to build a better AI assistant, stop tuning the model and start building a better environment for it to live in.