The Most Systematic 102‑Page Review of Agent Harnesses

Introduction

The rapid progress of large language models (LLMs) in code generation has shifted the focus from code as a final output to code as the operational foundation—called the Agent Harness—of autonomous systems.

Code as Agent Harness

The authors define Code as Agent Harness as code that is executable , inspectable , and stateful , making it the optimal medium to connect model reasoning, environment interaction, and persistent state.

Three‑Layer Architecture

Harness Interface (Interface Layer)

Code serves as a universal interface for agents, supporting three dimensions:

Reasoning : Externalizing intermediate computation to executables or symbolic solvers (Program‑Delegated Reasoning, Formal Verification, Iterative Code‑Grounded Reasoning).

Acting : Translating high‑level intent into executable policies (Grounded Skill Selection, Programmatic Policy Generation, Lifelong Code‑Based Agents).

Environment Modeling : Representing environments with executable artifacts (Structured World Representations, Execution‑Trace World Modeling, Code‑Grounded Evaluation Environments, Verifiable Environment Construction).

Harness Mechanisms (Mechanism Layer)

Ensures reliable long‑term execution through five interacting dimensions:

Planning : Four paradigms—Linear Decomposition, Structure‑Grounded Planning, Search‑Based Planning, Orchestration‑Based Planning.

Memory & Context Engineering : Five memory types—Working, Semantic, Experiential, Long‑Term, Multi‑Agent—plus context compaction and state offloading.

Tool Use : Four categories—Function‑Oriented, Environment‑Interaction, Verification‑Driven, Workflow‑Orchestration.

Harness Control : The Plan‑Execute‑Verify (PEV) loop with components—Plan as Contract, Sandboxed Execution, Permissioned State Transition, Deterministic Verification.

Agentic Harness Engineering (AHE) : Deep telemetry collects trace‑level data; an Evolution Agent diagnoses failures and proposes component revisions, which are evaluated and promoted via Governed Harness Mutation.

Scaling the Harness (Scaling Layer)

When task complexity exceeds a single agent's capacity, multi‑agent systems (MAS) share executable code artifacts, using execution feedback as objective convergence signals. Design space includes:

Specialized functional roles (Manager, Planner, Coder, Reviewer, Tester, Verifier).

Interaction modes (Collaborative Synthesis, Critique & Repair, Adversarial Validation, Reasoning Debate).

Workflow topologies (Waterfall, Agile/Iterative, Hierarchical, Star, Dynamic DAG).

Conclusion

The review serves as a detailed roadmap for researchers building coding agents, computer‑use agents, embodied AI, or scientific discovery agents.

Code as Agent Harness: Toward Executable, Verifiable, and Stateful Agent Systems
https://arxiv.org/pdf/2605.18747
https://github.com/YennNing/Awesome-Code-as-Agent-Harness-Papers