HarnessPro introduces a label‑free recursive self‑improvement framework that evolves agent harnesses to boost task performance while lowering execution cost through Pareto‑based optimization. Candidate harnesses are assessed solely with LLM‑generated unit tests, eliminating the need for ground‑truth solutions or official test suites. Specialized agents inspect execution trajectories, devise updates weighted toward different performance‑cost objectives, and materialize those updates as new candidate harnesses. Instead of outputting a single solution, HarnessPro employs a normalized weighted squared‑distance selector to produce three Pareto‑optimal variants—a cost‑oriented “lite” harness, a balanced “medium” harness, and a performance‑oriented “heavy” harness—tailored to distinct deployment budgets. Selected harnesses are fed into the next iteration, where fresh execution feedback drives further refinement. The method was evaluated on Terminal‑Bench 2, SWE‑Bench Pro, and ProgramBench using GPT‑5.5 paired with Codex and, separately, GLM‑5.2 paired with Claude Code. In the GPT‑5.5/Codex setting, the heavy mode of HarnessPro delivered an average pass‑rate improvement of 2.0 percentage points and an average per‑task cost reduction of 26.1% relative to the baseline RHO approach, with analogous gains observed when using GLM‑5.2/Claude Code. These results demonstrate that HarnessPro can simultaneously enhance effectiveness and efficiency across diverse coding benchmarks.