Harness Engineering: Anatomy, Architecture, and Evolution of Coding Agents

Abstract

Abstract. An agent is a model plus a harness: the runtime that couples an LLM to the world through a loop, tools, context management, safety controls, orchestration, and extension surfaces. This paper defines the harness, maps its seven canonical subsystems with the minimal and maximal implementation of each, and dissects eleven production coding harnesses along those subsystems. It does not benchmark or rank them; it describes and compares how the systems are built. Two absences and three extensibility standards survive a threefold corpus expansion across roughly four million lines of Python, TypeScript, and Rust.

Index Terms—coding agents, harness engineering, source-code study, MCP, SKILL.md, ACP

1 Introduction

A coding agent is not a model. It is a model plus a harness: the program that supplies the loop, tools, context, safety controls, orchestration, and extension surfaces around the weights. Harness engineering names the discipline of designing that runtime, and in the first half of 2026 it moved from a craft to a platform. Vendor SDKs shipped harnesses, marketplaces appeared, switching-cost tooling arrived, and enterprise governance started to treat the harness as the addressable unit. The corpus this paper studies is eleven production systems, read at source level rather than benchmarked.

The paper's most useful contribution is a boundary. The model is frozen inside the product; everything else can be rewritten, swapped, or governed. That boundary is where the design decisions live, and it is what this note reproduces in compact form.

Fig. 1. An agent is a model plus a harness of seven subsystems. The harness owns everything outside the weights.

2 The seven subsystems

Each subsystem has a minimal and a maximal implementation. The minimal column is what a credible harness must supply; the maximal column is what production systems added as the runtime matured. The gap between the two is where most of the source code lives.

Table 1 Minimal and maximal implementations of the seven harness subsystems.

Subsystem	Minimal implementation	Maximal implementation
Loop	Read, dispatch one tool, append result, repeat	Streamed plans, parallel tools, resumable checkpoints, budget guards
Tools	Shell plus a file editor	Sandboxed executors, LSP-aware edits, browser and search tools, typed schemas
Context	System prompt plus recent turns	Retrieval, compaction, memory files, per-task context assembly
Safety	Confirmation before writes	Permission modes, approval queues, secret redaction, audit trails
Orchestration	One process, one task	Subagents, parallel worktrees, queues, retries, cancellation
Extensions	None	MCP servers, skills, hooks, plugins, marketplaces
LLM integration	One provider, one key	Multi-provider routing, streaming, retries, usage accounting

3 Extensibility standards

Three standards appear across the corpus often enough to count as de facto interfaces. MCP wins the tool boundary; SKILL.md wins the instruction boundary; ACP is emerging at the agent-to-agent boundary. Adoption is uneven, and the paper reports it as counts out of eleven systems rather than as a ranking.

Fig. 2. Extensibility standards across eleven coding harnesses. Counts include partial support, greyed in the paper's original figure.

4 From tool to platform

The second half of 2026 is the convergence story. Codex adopts Claude Code's hook vocabulary; OpenHands reads its plugin format. Behavioral policy migrates from prompt prose into configuration, and the harness becomes an SDK with a marketplace attached. The paper's longitudinal sample of the same eight harnesses across a quarter shows the same drift in three of the eight systems.

Algorithm 1Harness selection under equal models

One decision procedure the paper recommends for teams choosing a harness.

1 Input: task set T, fixed model M, candidate harnesses H

2 for each h in H:

3 score[h] = held-out score of h running M on T

4 cost[h] = tokens + wall clock + human approvals

5 choose h maximizing score[h] subject to the cost budget

6 re-run the loop when M changes; never freeze the harness choice

Fig. 3. Convergence becomes imitation: Codex reads Claude Code's hook vocabulary while OpenHands reads its plugin format.

13 Cross-cutting observations

29 Design patterns

18 Recommendations

90 Minimum viable scaffold lines of code

Counts stated in the abstract of the study.

The corpus covers Claude Code, Codex CLI, Gemini CLI, Mistral Vibe, OpenHands, Aider, Mini-SWE-Agent, Hermes, Pi, OpenCode, and OpenClaw.
Omnigent is analyzed separately as a meta-harness that orchestrates the other ten behind one API.
Adoption counts include partial support and are reported as fractions of eleven systems.