Agent Harness
An execution framework that connects large language models with tools, responsible for task orchestration, state management, and error handling. It serves as the intermediary layer that coordinates the interaction between an AI model's reasoning capabilities and the external tools or services it invokes during task execution.
Core Facts
Agent Harness refers to the entire engineering system wrapped around a large language model that is responsible for coordinating the model's interactions with the external world.
A Harness is the collection of tools and runtime environment that an agent relies on, and is the underlying infrastructure that enables AI to perform operations.
The Agent Harness does not replace the reasoning capability of the core model, but instead builds a runtime management shell around the model, responsible for cross-cutting concerns such as tool invocation, state management, and error recovery.
Timeline (last 90 days)
In an Agent, the large model is only responsible for receiving messages and outputting messages; what actually performs operations such as file reading and command execution is the Agent Harness.
The accurate meaning of "the harness is dead" is that fancy harnesses can no longer add coding intelligence to models, not that harnesses themselves disappear entirely.
The harness is moving toward a higher level of abstraction: the previous generation combined a bunch of LLM calls, while the new generation treats the Agent as the core building block.
Although model evaluation scores are little affected by the harness, the cost per task differs significantly across harnesses.
Katz points out that there are three things in the harness that models cannot absorb: infrastructure and routing, trust and safety, and efficiency (cost optimization).
Jediah Katz defines Agent Harness as the system layer that "wraps the model, steers the model, and connects it to the world," divided into the Runtime layer and the Context orchestration layer.
MCP only addresses how capabilities are described and invoked, not permission narrowing, auditing, or cost control, which remain the harness's own responsibility.
Notion Auto involves a foundational rebuild of its Agents, centered on rebuilding the Agent Harness.
Hermes is essentially a "harness" (execution framework/shell) that does not itself produce intelligence but is responsible for organizing the complete closed loop of interaction with the model—including loop logic, skill invocation, system prompts, tool calls, and so on.
Optimizing the harness for a vertical domain can significantly improve agent reliability and efficiency, which is also why it is harder to customize than context.
40 more timeline events
All Facts (20)
Agent Harness refers to the entire engineering system wrapped around a large language model that is responsible for coordinating the model's interactions with the external world.
80%VerifiedA Harness is the collection of tools and runtime environment that an agent relies on, and is the underlying infrastructure that enables AI to perform operations.
80%VerifiedThe Agent Harness does not replace the reasoning capability of the core model, but instead builds a runtime management shell around the model, responsible for cross-cutting concerns such as tool invocation, state management, and error recovery.
80%VerifiedThe technical foundation of the Agents API is powered by the Codex harness.
75%VerifiedAgentConnect supports automatic triggering of agent tasks from multiple sources including conversations, Pull Requests, Issues, Webhooks, and scheduled tasks.
75%VerifiedThe concept of Harness originates from Test Harness in software engineering, and in the AI Agent context it is extended to mean the systematic engineering that enables an Agent to run continuously, cyclically, and in a controlled manner
75%VerifiedAgent Harness is an engineering constraint framework built around agents, with the concept borrowed from the wiring harness system in automotive engineering.
75%VerifiedMany agent frameworks are deliberately designed to isolate final evaluation data from the optimization loop, so agents can only see training and validation signals and cannot directly access the test set
65%VerifiedOn first run, Harness prompts the user to enter a DeepSeek API key, but users can connect to other providers in the settings, including GPT-series models via OpenCode and local Ollama.
65%Verified工具调用（Function Calling/Tool Use）是Agent系统的核心能力，允许模型动态调用搜索引擎、代码解释器、数据库查询、API接口等外部资源
65%UnverifiedSuperpowers本质上是一个专为Coding场景设计的轻量级Agent Harness实现
85%UnverifiedIn an Agent, the large model is only responsible for receiving messages and outputting messages; what actually performs operations such as file reading and command execution is the Agent Harness.
50%UnverifiedThe accurate meaning of "the harness is dead" is that fancy harnesses can no longer add coding intelligence to models, not that harnesses themselves disappear entirely.
50%UnverifiedThe harness is moving toward a higher level of abstraction: the previous generation combined a bunch of LLM calls, while the new generation treats the Agent as the core building block.
50%UnverifiedKatz points out that there are three things in the harness that models cannot absorb: infrastructure and routing, trust and safety, and efficiency (cost optimization).
50%UnverifiedAlthough model evaluation scores are little affected by the harness, the cost per task differs significantly across harnesses.
50%UnverifiedJediah Katz defines Agent Harness as the system layer that "wraps the model, steers the model, and connects it to the world," divided into the Runtime layer and the Context orchestration layer.
50%UnverifiedMCP only addresses how capabilities are described and invoked, not permission narrowing, auditing, or cost control, which remain the harness's own responsibility.
50%UnverifiedNotion Auto involves a foundational rebuild of its Agents, centered on rebuilding the Agent Harness.
50%UnverifiedHermes is essentially a "harness" (execution framework/shell) that does not itself produce intelligence but is responsible for organizing the complete closed loop of interaction with the model—including loop logic, skill invocation, system prompts, tool calls, and so on.
50%