PapersSelected

Google paper proposes GUARD for self-improving agent harnesses

Why we picked it

The non-obvious insight is that accept/reject gating on regression checks, not bigger rewrites, is what keeps self-improving harnesses stable — aggressive rewriters matched its scores but tripled memory pollution. Developers building agent frameworks or coding assistants should study the self-check and rule-pinning mechanism, as it can be retrofitted into existing harnesses. The token efficiency win (54.5% fewer per task versus Meta-Harness) also matters for teams managing inference costs in production agents.

A new Google research paper introduces GUARD, a method for building self-improving agent harnesses that avoids common failure modes such as local overfitting, catastrophic forgetting, and memory pollution. The system keeps the base model frozen and lets the agent edit only its own prompts, skills, tool rules, and memory. An edit is accepted only when its measured gains outweigh the behaviors it breaks.

Body

GUARD validates changes by splitting each task into self-checks executed before the task runs. Every improvement round grades trajectories against these checks, pins rules the agent already handles reliably, and prioritises newly broken behaviors when deciding whether to accept an edit. At test time, a single frozen harness runs in one pass with no further self-modification. On 43 hosted SDK tasks, GUARD reached 80.80% accuracy, 8.89 points above the starting harness (p = 0.008). It beat Meta-Harness by 6.28 points while using 54.5% fewer tokens per task. On procedural memory tasks, it achieved 90.1% update correctness with zero memory pollution, whereas aggressive rewriters scored similarly but tripled pollution.

The work targets a recurring pain point in agent engineering: harnesses that improve themselves often degrade other capabilities or accumulate corrupted memory. GUARD offers a guardrail-first recipe that developers building autonomous coding or tool-use agents can adopt, with the paper and a Chat with Paper interface available via arXiv (2611.23476) and dair.ai.

Original · elvis

Banger paper from Google on self-improving agent harnesses. (bookmark it) It addresses common issues in self-improving agent harnesses like local overfitting, catastrophic forgetting, and memory pollution. They propose GUARD, which keeps the base model frozen and lets the agent edit its own prompts, skills, tool rules, and memory. An edit is accepted only when its gains outweigh the behaviors it breaks. - On 43 hosted SDK tasks, GUARD reached 80.80% accuracy, 8.89 points over the starting harness (p = 0.008). - It beat Meta-Harness by 6.28 points and used 54.5% fewer tokens per task. - On procedural memory tasks, it led update correctness at 90.1% with zero memory pollution. Aggressive rewriters scored similarly and tripled pollution. How does it know what broke? Each task is split into self-checks before execution. Every round grades trajectories against them, pins rules the agent already handles, and ranks newly broken behaviors first. At test time, one frozen harness runs in a single pass. arxiv.org/abs/2611.23476 Chat with Paper: academy.dair.ai/papers/guard-s…📊 14 ⚡

Industry Coordinates & Source Verification

This report is categorized under AITOP's "Papers" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include GUARD, agents, Google, papers.

Source attribution: curated from elvis (@omarsar0).