長時間動作アプリケーション開発向けハーネス設計Harness design for long-running application development

In agentic coding, the model is only part of the equation. The surrounding software that feeds a model context, exposes tools, manages state, and decides when to stop or continue, often called the harness, increasingly determines how well an AI system performs on real engineering work. Anthropic's engineering team has detailed how careful harness design helped push Claude further in two demanding areas: frontend design and long-running autonomous software engineering. The topic matters because as tasks stretch from seconds to hours, the difference between a capable model and a useful system comes down to the scaffolding around it.

A harness is the orchestration layer that sits between a user request and the model. It assembles the prompt, manages the context window, registers the tools the agent can call, captures the results of those calls, and loops until a goal is reached. For short interactions, a thin harness is often enough. For tasks that span many steps, the harness must handle accumulating context, recover from errors, and keep the agent oriented toward the original objective without drifting or repeating work. Anthropic frames this as a design discipline in its own right, distinct from training the underlying model.

Long-running tasks expose problems that short prompts never reveal. As an agent works for an extended period, its context window fills with tool outputs, intermediate reasoning, and prior actions. Without management, important early instructions can be crowded out, and the agent may lose track of what it has already tried. Techniques to address this typically include compacting or summarizing earlier steps, persisting state to external files rather than relying solely on the context window, and structuring work so the agent can resume cleanly after an interruption. The goal is to let an agent operate for hours while remaining coherent, a prerequisite for autonomous software engineering where a single task might involve reading a codebase, making changes across many files, running tests, and iterating on failures.

Frontend design is a useful proving ground because it combines code generation with visual judgment. Producing a working interface is not only about syntactically correct code; it involves layout, styling, responsiveness, and aesthetic choices that are hard to specify precisely. A harness can improve results here by giving the agent feedback loops, for example by letting it render output, inspect the result, and revise. Anthropic indicates that pushing Claude on frontend work involved refining how the agent perceives and iterates on what it builds, rather than expecting a correct answer in a single pass. This reflects a broader pattern in agentic systems, where the ability to verify and self-correct often matters more than raw first-attempt accuracy.

The work sits within a fast-moving industry context. Anthropic ships Claude Code, a command-line agent for software engineering, and has released the Claude Agent SDK and the Model Context Protocol, an open standard for connecting models to external tools and data sources. Competing efforts, including OpenAI's coding agents and a range of third-party tools such as Cursor, Aider, and Devin, are converging on similar concerns: how to give models reliable access to a development environment, how to manage long horizons of work, and how to evaluate progress. Benchmarks like SWE-bench, which measures whether an agent can resolve real GitHub issues, have helped standardize comparison, though they capture only part of what production systems require.

Several prerequisite concepts underpin this approach. Tool use, sometimes called function calling, lets a model invoke external operations such as editing files or running commands and then incorporate the results. Context management addresses the finite window of tokens a model can attend to at once. Evaluation and feedback loops give the agent signals about whether its work is succeeding. Effective harness design weaves these together so that capability already present in the model can be reliably elicited on hard, multi-step tasks.

The broader implication is that improving agent performance is not solely a matter of larger or newer models. Much of the gain appears to come from engineering the system around the model, an area that remains relatively young and where practices are still being established. For teams building on Claude or comparable systems, Anthropic's account suggests that investing in the harness, especially state management and verification for long-running work, is likely to yield meaningful returns alongside any underlying model improvements.