Google Researchers Introduce GUARD for Stable Self-Improving Agent Harnesses

Researchers from Google Cloud introduced GUARD, a grader-free closed-loop framework that lets autonomous AI agents edit their own prompts, skills, tool rules, and memory while keeping the base model frozen. GUARD decomposes each task into pre-execution self-checks and uses directional regret to accept an edit only when its gains outweigh the behaviors it breaks. On 43 hosted SDK tasks, GUARD reached 80.80% accuracy, 8.89 points above the starting harness, and used 54.5% fewer tokens per task.
Event version 1.
Key insights
Regret-guided edits
GUARD splits each task into pre-execution self-checks and grades trajectories against them each round. It pins rules the agent has mastered and ranks newly broken behaviors first, making regressions visible.
Soft-HURT gate
Parallel candidate edits are compared by critic-free group-relative fitness, and the Soft-HURT gate promotes an edit only when gains outweigh regressions on its gated buffer. Pooling two candidates in parallel cut rejected gate rounds by 64% against sequential search at equal budget.
Memory without pollution
On procedural memory tasks GUARD led task accuracy at 83.0% and update correctness at 90.1% with 0.000 procedural pollution. Aggressive rewriters such as Meta-Harness reached similar update scores but tripled pollution.
Hosted SDK results
On 43 test tasks GUARD reached 80.80% accuracy, 8.89 points over the starting harness (paired p = 0.008) and 6.28 points over Meta-Harness, while using 54.5% fewer tokens per task. At test time one frozen harness runs in a single pass.