RSI Diving into the Harness Engineering: Pi Makes the Agent Run Longer, DeepSeek Makes the Agent Adapt

According to Dynamic Insight Beating monitoring, recursive self-improvement is increasingly manifesting in a more tangible manner on Harness. Recently, the designs of Pi and DeepSeek Harnesses have precisely filled in two key pillars: one to allow the Agent to run for a sufficient period and another to make the Agent itself easily modifiable.

The latest Pi Harness V3 specification transforms the Agent into a resilient persistent runtime. Before model requests and tool invocations, the execution intent is recorded, and upon completion, the results and next-step state are written. In the event of a process crash, the system can determine where the task halted, which operations can be safely rerun, and which have side effects and cannot be re-executed. For the Agent to run long-term, it must not restart from scratch after each crash.

DeepSeek Harness complements the other half. Cordis has made the model adapter, tooling system, Session Log, and even the Agent Loop into composable components; the creative mode also allows the Agent to inspect the Cordis environment at runtime, dynamically define, load, and unload new components.

Pi enables the continuity of execution state, while DeepSeek allows the dynamic restructuring of the Agent's composition.

This aligns with Weng Li's direction discussed in July regarding the Harness and the recursive self-improvement judgment. In the near term, RSI may not necessarily begin with a direct rewrite of weights from the model; a more practical path is to first optimize the context and workflow, then delve into the Harness code, with even "how to optimize the Harness" itself becoming the target of optimization.

Now, the third piece is still missing: validation. Research on self-evolving Agents has already identified evaluation and security as core issues. An Agent can run long-term and modify itself, but if even the definition of "progress" can be altered together, it is easy to optimize for passing its own tests. Weng Li also particularly emphasizes that the verifier, tracer, and security boundaries should reside outside the self-modification loop.

Therefore, a true RSI closed loop must address at least three things: running long, being modifiable, and validating accurately. Pi is addressing the first piece, DeepSeek is addressing the second piece, and the third piece will determine whether this "evolution" truly results in enhanced capabilities or if it merely learns to pass its self-created exams.

Being able to modify oneself is just the beginning of evolution; proving oneself has genuinely become stronger is the most challenging final stretch of RSI.