Self-Modifying Agents, Runtime Security, and Smarter Skill Loading

Research

Meta's HyperAgents: AI That Rewrites Its Own Improvement Strategy

Every self-improving AI system has had a ceiling where the meta-agent's strategy was locked by human researchers. HyperAgents removes that ceiling by letting the meta-agent rewrite its own modification code, and the cross-domain transfer results are hard to ignore: trained on paper review and robotics, it still hits 0.630 on Olympiad math.

Analysis

DGM-HyperAgents Benchmark Breakdown: What the Numbers Actually Show

The concrete benchmark gains tell the real story: paper review jumped from 0.0 to 0.710, coding from 0.084 to 0.267, and robotics reward design from 0.060 to 0.372. If these results hold up to scrutiny, this is the most credible demonstration of open-ended self-improvement transfer we have seen.

Tools

Google ADK's SkillToolset: Build Agents That Write Their Own Skills at Runtime

Google's Agent Development Kit now supports a skill factory pattern where agents generate new expertise on demand instead of bloating the system prompt with thousands of tokens upfront. The progressive disclosure architecture (metadata at startup, full instructions only when needed) is a practical solution to a real scaling problem every agent builder hits.

Akeyless Ships Intent-Aware Runtime Security for AI Agents

As agents move from reading data to actually modifying systems and executing workflows, static OAuth tokens and role-based access controls are not built for the job. Akeyless's Runtime Authority enforces Zero Standing Privileges and blocks unsafe actions at the moment of execution, which is the right architectural layer to solve this.