Harness RSI Progress: Pi Enables Long-Run Agents, DeepSeek Allows Modification

Pi and DeepSeek advance Harness RSI by enabling persistent agent runtimes and dynamic component modification. However, accurate verification remains a critical missing element for true recursive self-improvement.

Woofun AI reports that recursive self-improvement in the Harness project is advancing through distinct technical contributions from Pi and DeepSeek. Pi's Harness V3 specification establishes agents as recoverable, persistent runtime systems that record intended actions before execution and save results and state afterward, allowing processes to resume safely after crashes without restarting from scratch. DeepSeek's Harness addresses structural flexibility by using Cordis to turn model adapters, tool systems, session logs, and agent loops into composable components, enabling agents to inspect their environment at runtime and dynamically load or unload new parts.

While Pi ensures execution continuity and DeepSeek facilitates dynamic reorganization, the third essential element—accurate verification—remains unresolved. Weng Li previously noted that true RSI requires optimizing contexts and workflows before modifying Harness code, with verification, tracing, and security boundaries placed outside the self-modification loop to prevent agents from merely optimizing to pass self-designed tests.