[RSI Begins to Land on Harness Engineering: Pi Enables Agents to Run Longer, DeepSeek Enables Agents to Adapt] According to monitoring by Dongcha Beating, recursive self-improvement is increasingly being concretely implemented on Harness. Recently, the two designs of Pi and DeepSeek Harness have precisely filled two critical foundational gaps: one enables agents to run long enough, and the other makes agents themselves easy to modify. Pi's latest Harness V3 specification designs agents as recoverable persistent runtimes. Before model requests and tool calls, execution intentions are recorded first, and after completion, results and next-step states are written. If a process crashes, the system can identify where the task stopped, which operations can be safely rerun, and which have side effects and cannot be executed again. For agents to run long-term, they must not restart from scratch after every crash. DeepSeek Harness addresses the other half. Cordis turns model adapters, tool systems, session logs, and even agent loops into composable components; the creation mode also allows agents to inspect the Cordis environment during runtime, temporarily define, load, and unload new components. Pi enables execution states to persist, while DeepSeek enables the structural composition of agents to dynamically reorganize. This aligns with the direction discussed by Weng Li in July regarding Harness and recursive self-improvement. In the near term, RSI may not start by directly rewriting model weights; a more practical route is to first optimize context and workflows, then delve into Harness code, and eventually make "how to optimize Harness" itself an optimization target. The third piece is still missing: accurate verification. Research on self-evolving agents has already listed evaluation and safety as core issues. Agents can run long-term and modify themselves, but if even the definition of "progress" can be altered, it becomes easy to optimize merely to pass their own tests. Weng Li also emphasized that verifiers, tracers, and safety boundaries should be placed outside the self-modification loop. Therefore, a true RSI closed loop must solve at least three things: running long, adapting effectively, and verifying accurately. Pi is addressing the first piece, DeepSeek is addressing the second, and the third piece determines whether this "evolution" ultimately leads to genuine capability growth or merely learning to pass its own exams. The ability to modify oneself is just the beginning of evolution; proving genuine improvement is the hardest final mile for RSI. [Original Link]