Externalization in LLM Agents: A Unified Review of Memory, Skills, Protocols and Harness Engineering

arxiv:2604.08224

Original source · Paper JSON

Harness · 2026-04-09

Cites (6) · Cited by (0)

Complete cleaned Markdown · Source Markdown including bibliography

Read in parts

1. Part 1
2. Part 2
3. Part 3
4. Part 4
5. Part 5
6. Part 6
7. Part 7
8. Part 8
9. Part 9
10. Part 10
11. Part 11

Contents

• Externalization in LLM Agents: A Unified Review of Memory, Skills, Protocols and Harness Engineering
• Abstract
• 1 Introduction
• Memory systems externalize state across time.
• Skill systems externalize procedural expertise.
• Protocols externalize interaction structure.
• 2 Background: From Weights to Context to Harness
• 2.1 Capability in Weights
• 2.2 Capability in Context
• 2.3 Capability through Infrastructure
• 2.4 Externalization as the Transition Logic
• 3 Externalized State: Memory
• 3.1 What Is Externalized: The Content of State
• Working context.
• Episodic experience.
• Semantic knowledge.
• Personalized memory.
• 3.2 How It Is Externalized: Memory Architectures
• 3.2.1 Monolithic Context
• 3.2.2 Context with Retrieval Storage
• 3.2.3 Hierarchical Memory and Orchestration
• 3.2.4 Adaptive Memory Systems
• 3.3 Memory Demands of the Harness Eras
• 3.4 Memory as Cognitive Artifact
• 4 Externalized Expertise: Skills
• 4.1 What is Externalized: Procedural Expertise
• 4.1.1 Operational Procedure
• 4.1.2 Decision Heuristics
• 4.1.3 Normative Constraints
• 4.2 From Execution Primitives to Capability Packages
• 4.2.1 Stage 1: Atomic Execution Primitives
• 4.2.2 Stage 2: Large-scale Primitive Selection
• 4.2.3 Stage 3: Skill as Packaged Expertise
• 4.3 How Skills Are Externalized
• 4.3.1 Specification
• 4.3.2 Discovery
• 4.3.3 Progressive Disclosure
• 4.3.4 Execution Binding
• 4.3.5 Composition
• 4.4 Skill Acquisition and Evolution
• Authored.
• Distilled.
• Discovered.
• Composed.
• 4.5 Boundary Conditions
• Semantic alignment.
• Portability and staleness.
• Unsafe composition.
• Context-dependent degradation.
• 4.6 Skills in the Harness
• Conditioning on memory.
• Binding through protocols.
• Runtime governance.
• Lifecycle feedback.
• 4.7 Skill as Cognitive Artifact
• 5 Externalized Interaction: Protocols
• Invocation grammar.
• Lifecycle semantics.
• Permission and trust boundaries.
• Discovery metadata.
• 5.1 Why Protocols Matter
• Unified interaction standards.
• Improved security, governance, and auditability.
• Reduced vendor dependence.
• 5.2 Agent Protocol Survey
• 5.2.1 Agent-Tool Protocols
• 5.2.2 Agent-Agent Protocols
• 5.2.3 Agent-User Protocols
• 5.2.4 Other Protocols
• 5.3 Agent Protocol in Harness Engineering
• 5.3.1 Intent Capture and Normalization
• 5.3.2 Capability Discovery and Tool Description
• 5.3.3 Session and Lifecycle Management
• 5.4 Protocol as Cognitive Artifact
• 6 Unified Externalization: Harness Engineering
• 6.1 What is a Harness?
• 6.2 Analytical Dimensions of Harness Design
• 6.2.1 Agent Loop and Control Flow
• 6.2.2 Sandboxing and Execution Isolation
• 6.2.3 Human Oversight and Approval Gates
• 6.2.4 Observability and Structured Feedback
• 6.2.5 Configuration, Permissions, and Policy Encoding
• 6.2.6 Context Budget Management
• 6.3 Harness in Practice: Contemporary Agent Systems
• Loop and control flow.
• Sandboxing.
• Human oversight.
• Observability.
• Configuration and governance.
• Context budget.
• 6.4 Harness as Cognitive Environment
• 7 Cross-Cutting Analysis
• 7.1 Module Interaction Map
• Memory to skill: experience distillation.
• Skill to memory: execution recording.
• Skill to protocol: capability invocation.
• Protocol to skill: capability generation.
• Memory to protocol: strategy selection.
• Protocol to memory: result assimilation.
• System-level dynamics.
• 7.2 The LLM Input/Output Perspective
• Memory as contextual input.
• Skills as instructional input.
• Protocols as action schema.
• 7.3 Parametric vs. Externalized: The Trade-off Space
• Update frequency and temporal decay.
• Reusability and multi-agent portability.
• Auditability, governance, and alignment.
• Latency, simplicity, and context burden.
• 8 Future Discussion
• 8.1 The Expanding Frontier
• Planning and goal management.
• Evaluation and verification.
• Orchestration logic itself.
• Multi-modal externalization.
• 8.2 From Digital Agents to Embodied Externalization
• The monolithic starting point.
• Decomposition: the cerebrum–cerebellum split.
• Why the parallel matters.
• Open challenges.
• 8.3 Toward Self-Evolving Harnesses
• 8.4 Costs, Risks, and Governance
• Cognitive overhead.
• Security and integrity risks.
• Governance as infrastructure.
• 8.5 From Private Scaffolding to Shared Infrastructure
• Shared artifacts.
• Division of labor and collective learning.
• Institutionalization and its tensions.
• 8.6 Measuring Externalization
• 9 Conclusion