ForgeCode is a code execution harness developed by Anthropic for post-training large language models, specifically designed to enhance code execution and terminal command capabilities. The harness gained prominence through its integration into the training pipeline for Claude Opus 4.6, achieving a 79.8% accuracy rate on Terminal-Bench 2.0 benchmarks, establishing it as Anthropic's optimized harness choice for this model series.

ForgeCode functions as a structured environment for executing and validating code generated by language models. A code execution harness provides a controlled sandbox in which generated code can be safely executed, tested, and evaluated against defined benchmarks. This approach enables models to receive immediate feedback on code correctness during the post-training phase, facilitating supervised fine-tuning and reinforcement learning from human feedback (RLHF) that improves code generation capabilities ¹⁾.

Post-training techniques such as instruction tuning and reinforcement learning rely on high-quality feedback signals to guide model behavior. By incorporating ForgeCode into Claude Opus 4.6's training process, Anthropic created a direct feedback mechanism for code execution tasks. The harness evaluates generated code against test cases and benchmarks, enabling the model to learn from successes and failures in a systematic manner. This aligns with established practices in instruction fine-tuning, where models are trained to follow specific task requirements with measurable outcomes ²⁾.

Terminal-Bench 2.0 represents a standardized evaluation benchmark for assessing code generation and command execution capabilities in language models. ForgeCode's achievement of 79.8% accuracy on this benchmark demonstrates its effectiveness in enabling Claude Opus 4.6 to generate valid terminal commands and executable code sequences. This performance metric reflects both the quality of the generated code and the model's understanding of command-line semantics, file systems, and shell scripting patterns. Terminal command generation presents particular challenges due to environment-specific variations, security considerations, and the need to understand contextual requirements for execution.

Code execution harnesses typically implement several key components: a sandbox environment for safe execution, a validation framework to check output correctness, test case management, and feedback collection systems ³⁾. ForgeCode's design as Anthropic's optimized choice suggests careful engineering of these components to maximize both safety and learning signal quality. The harness likely includes error detection, output comparison mechanisms, and integration points with reinforcement learning systems that reward correct code generation and penalize unsafe or incorrect commands.

The integration of ForgeCode into Claude Opus 4.6 enhances the model's utility for software development tasks, DevOps automation, system administration, and technical problem-solving. Models trained with robust code execution harnesses demonstrate improved performance across multiple code-related benchmarks and real-world applications. The 79.8% Terminal-Bench 2.0 accuracy enables practical deployment for code generation, script writing, and command-line instruction tasks where correctness is critical. This capability supports developers in automating repetitive tasks, generating boilerplate code, and exploring command-line options ⁴⁾.

• Forge
• Capy
• Copilot