The Last Harness You'll Ever Build

AI agents are increasingly deployed on complex, domain-specific workflows – navigating enterprise web applications that require dozens of clicks and form fills, orchestrating multi-step research pipelines that span search, extraction, and synthesis, automating code review across unfamiliar repositories, and handling customer escalations that demand nuanced domain knowledge. $\mathbf{\text{Each}}\mathbf{\text{new}}\mathbf{\text{task}}\mathbf{\text{domain}}\mathbf{\text{requires}}\mathbf{\text{painstaking,}}\mathbf{\text{expert-driven}}\mathbf{\text{harness}}\mathbf{\text{engineering}}$: designing the prompts, tools, orchestration logic, and evaluation criteria that make a foundation model effective. We present a two-level framework that automates this process. At the first level, the $\mathbf{\text{Harness}}\mathbf{\text{Evolution}}\mathbf{\text{Loop}}$ optimizes a worker agent’s harness $\mathcal{H}$ for a single task: a Worker Agent $W_{\mathcal{H}}$ executes the task, an Evaluator Agent $V$ adversarially diagnoses failures and scores performance, and an Evolution Agent $E$ modifies the harness based on the full history of prior attempts. At the second level, the $\mathbf{\text{Meta-Evolution}}\mathbf{\text{Loop}}$ optimizes the evolution blueprint $\mathrm{\Lambda }=(W_{\mathcal{H}},{\mathcal{H}}^{(0)},V,E)$ itself across diverse tasks, $\mathbf{\text{learning}}\mathbf{\text{a}}\mathbf{\text{blueprint}}{\mathrm{\Lambda }}^{(\mathbf{\text{best}})}\mathbf{\text{that}}\mathbf{\text{enables}}\mathbf{\text{rapid}}\mathbf{\text{harness}}\mathbf{\text{convergence}}\mathbf{\text{on}}\mathbf{\text{any}}\mathbf{\text{new}}\mathbf{\text{task}}\mathbf{\text{–}}\mathbf{\text{so}}\mathbf{\text{that}}\mathbf{\text{adapting}}\mathbf{\text{an}}\mathbf{\text{agent}}\mathbf{\text{to}}\mathbf{\text{a}}\mathbf{\text{novel}}\mathbf{\text{domain}}\mathbf{\text{requires}}\mathbf{\text{no}}\mathbf{\text{human}}\mathbf{\text{harness}}\mathbf{\text{engineering}}\mathbf{\text{at}}\mathbf{\text{all.}}$ We formalize the correspondence to meta-learning and present both algorithms. The framework $\mathbf{\text{shifts}}\mathbf{\text{manual}}\mathbf{\text{harness}}\mathbf{\text{engineering}}\mathbf{\text{into}}\mathbf{\text{automated}}\mathbf{\text{harness}}\mathbf{\text{engineering}}$, and takes one step further – $\mathbf{\text{automating}}\mathbf{\text{the}}\mathbf{\text{design}}\mathbf{\text{of}}\mathbf{\text{the}}\mathbf{\text{automation}}\mathbf{\text{itself}}$.