Recursive self-improvement (RSI) involves AI systems enhancing their intelligence through feedback loops. Yudkowsky (2008) defined RSI as an AI improving its cognitive machinery, which may include rewriting weights or optimizing deployment systems. Modern AI advancements, like those from Anthropic and Open AI, show accelerated RSI via harness engineering. A harness is a system orchestrating model execution, tool calls, context management, and evaluation. It extends beyond prompts to runtime software design, enabling generalization through workflow automation, persistent memory, and sub-agent delegation.

Workflow automation involves goal-oriented loops (plan, execute, observe, improve). Karpathy’s autoresearch repo exemplifies this, with models iterating via agent runtime rather than static prompts. File system management is critical for long-horizon tasks, as models must handle artifacts beyond context windows. Tools like bash commands and file operations (read/write/edit) become foundational skills.

Sub-agents and backend jobs allow parallel task execution. A harness can spawn sub-agents for experiments or subtasks, requiring a process manager to track logs and results. This mirrors OS-like encapsulation, simplifying interfaces while managing complexity. Coding agents (e.g., Claude Code, Codex) demonstrate this with loops involving tool calls, file systems, and external context. Harness optimization evolves from prompts to workflow code, with future RSI likely relying on meta-methodologies rather than direct model weight changes.

Harness engineering is pivotal for RSI’s near-term path, enabling auto-research and sustainable systems. Context management and standardization (e.g., tool interfaces) will be key. The trajectory mirrors prompt engineering’s shift from manual tricks to structured contexts, emphasizing generalization over specificity.