Skip to content

Harness Engineering Mapping

Map the four harness elements onto the 5-layer model and clarify what harness covers and what it doesn't.

About This Document

NOTE

"Harness Engineering" emerged as a notable framing in 2025–2026 for organizing the implementation mechanisms used to operate an LLM. Since this site is a map for designing LLM agents, this page makes the relationship between the two explicit.

For readers who found this site while searching for harness engineering: this page maps the four harness elements onto the site's 5-layer model and shows why the areas harness does not cover (Skills layer, part of Doctrine layer, the Why behind each prescription) still require a separate framework.

TIP

In three lines

  • Harness is the mechanism for operating; the 5-layer model is the map for designing. They live at different abstraction levels.
  • The four harness elements map onto MCP, Memory, Agent, and Doctrine (defensive aspect only).
  • The Skills layer and the offensive side of the Doctrine layer (normative strength declarations) are not part of harness, so this site fills that gap.

What Is Harness Engineering

NOTE

In this document, "harness engineering" refers to the implementation mechanism composed of the following four elements.

ElementDescription
Action (tool integration)Connection points for accessing external APIs, databases, file systems, browsers, etc.
Context (memory)A mechanism that retains past context, business background, and agent action history, and hands it back to the LLM when needed
Guardrails (safety controls)Safety devices (sandboxes, etc.) that prevent leakage of confidential information and uncontrolled system damage
Orchestration (loop control)A continuous loop that decomposes tasks, lets the LLM think, executes, evaluates results, and decides the next action

The word "harness" — like the harness used in rocketry or climbing — carries the metaphor of a fixture or restraint and is used as the implementation mechanism beneath higher-level methodologies such as Agent Engineering or Context Engineering.

Mapping to the 5-Layer Model

Mapping Table

Harness Element5-Layer ModelCorrespondence
Action (tool integration)MCPConnection points for external systems. Protocol layer.
Context (memory)MemoryPersistent memory and relationships. Knowledge Graph, etc.
Guardrails (safety controls)Doctrine (defensive aspect only)Constraints, prohibitions, sandboxing. Does not include normative strength declarations (MUST/SHOULD/MAY) or offensive design guidance.
Orchestration (loop control)AgentThe locus of task decomposition, execution, and evaluation loops.
❌ no counterpartSkillsStatic knowledge, guidelines, progressive disclosure. Absent in the harness vocabulary.

Three Areas Harness Does Not Cover

1. The Skills Layer (Reference Model for Static Knowledge)

The four harness elements lack the concept of "how to structure static knowledge and when to invoke it." The Skills layer covers:

  • The SKILL.md format and progressive disclosure
  • Auto-triggering based on description matching
  • The choice between Skills and MCP (see Sub-agent vs Skills)
  • Composition across multiple Skills

These are neither Context (memory) nor Action (tools); they are judgment criteria conditionally injected into the LLM's immediate context and require an independent layer.

IMPORTANT

Squeezing Skills into Context causes token bloat and Priority Saturation. The design hinges on expanding Skills only when invoked — explored in detail in concepts/03-architecture.

2. The Offensive Side of the Doctrine Layer

The Guardrails element of harness is confined to defensive functions: leak prevention and runaway prevention. The Doctrine layer also covers:

  • Purpose declarations (what the agent exists for)
  • Judgment criteria (priority order when trade-offs arise)
  • Normative strength ladder (MUST / SHOULD / MAY, per RFC 2119)
  • Role boundaries (what the agent addresses and what it does not)

These are "offensive design guidance" and have no counterpart in the harness vocabulary. See concepts/07-doctrine-and-intent for detail.

3. The Why Behind Each Prescription

Harness prescribes "give it memory" and "wrap it in a loop," but does not explain why those prescriptions are necessary:

  • Why externalize Context? → Context Rot, Lost in the Middle
  • Why re-inject instructions in a loop? → Instruction Decay, Priority Saturation
  • Why sandbox at all? → Hallucination, Sycophancy

These structural constraints are the subject of the sister site, understanding-llm-through-claude-code.

"Is Harness Enough?" Decision Table

Determine whether harness alone covers your question:

Reader's QuestionHarness Enough?Where to Look Next
"I want to give the LLM tools"✅ Yes
"I need to design memory"⚠️ PartialWhy (Context Rot, Lost in the Middle) → understanding-llm
"Should this go in Skills or MCP?"❌ Noconcepts/03-architecture, skills/what-is-skills
"Criteria for splitting sub-agents"❌ Noagents/subagent-vs-skill, agents/subagent-quality-gate
"What to write as MUST vs SHOULD"❌ Noconcepts/07-doctrine-and-intent
"Instructions decay over long tasks"⚠️ Symptomatic onlyWhy (Instruction Decay) → understanding-llm
"Granularity of guardrails"⚠️ Defensive onlyDoctrine (offensive normativity)
"Coordination across multiple MCPs and Skills"❌ Nostrategy/composition-patterns

Vocabulary Hierarchy — Harness Is a Mechanism, Engineering Is a Methodology

  • Harness = a noun, a thing. Like a rocket or climbing harness, it is a fixture metaphor → persists as mechanism.
  • X Engineering = methodology label. Agent Engineering, Context Engineering, etc. swap in and out as the umbrella term → disposable.

The 5-layer model and the 8 problems on the sister site are designed as vocabulary-independent abstractions. Whenever a new methodology label trends, a mapping page like this one can be added without disturbing the core model.

Reframing With Three Verbs

VerbGoalOutputTime Horizon
OperateControl the LLM to complete the taskA working agent (runtime)Today
DesignBuild reusable structure and judgment criteriaA design map (5-layer model + Doctrine)Sustained next year
UnderstandGrasp why the constraint existsThe why bookshelf (8 problems)Invariant

IMPORTANT

The three are complementary at different layers, not interchangeable. Harness handles operate, this site handles design, and the sister site handles understand.

🔗 Go Deeper: Why Each Harness Element Is Necessary

This page covers the structural correspondence (What) between harness and the 5-layer model. For why each harness element is necessary in terms of LLM structural constraints, see the sister site.


Previous: Local LLM Workspace MappingNext: Permission vs. Authority

Released under the MIT License.