New research reveals a surprising truth: when AI agents try to improve themselves by updating their own instructions and tools, the smartest models aren't always the best teachers—and mid-tier models benefit most from the upgrades.
The Self-Evolving Agent Experiment
Researchers tested whether large language models could improve their own performance by editing their external "harnesses"—the prompts, skills, memories, and tools that guide how they work. Think of it like an AI agent rewriting its own instruction manual after learning from mistakes.
The findings challenge conventional wisdom. When models like Claude Opus 4.6 and smaller models like Qwen3.5-9B generated updates to agent harnesses, the improvements were surprisingly similar. A 9-billion parameter model's updates worked nearly as well as those from frontier models—suggesting that creating useful updates doesn't require cutting-edge intelligence.
The Mid-Tier Sweet Spot
But here's where it gets interesting: not all models benefit equally from better instructions. Weak models couldn't follow the improved guidance. Strong models didn't need much help. Mid-tier models—the workhorses most businesses actually deploy—saw the biggest gains.
The researchers identified two failure modes in weaker models: they either failed to recognise when to use updated tools, or activated them but couldn't follow multi-step instructions faithfully. This matters because it suggests where to invest training effort: not in making agents better self-improvers, but in making them better at using improvements.
What This Means for Learners
If you're building AI agents or exploring agentic workflows, this research offers practical guidance. Focus your capability budget on the agent doing the work, not the one trying to improve it. A mid-tier model with well-designed harnesses often outperforms a frontier model with poorly structured tools.
Understanding how to structure prompts, skills, and memory systems—what the paper calls "harnesses"—is now a core AI engineering skill. The research also highlights the importance of long-horizon instruction following: agents need to reliably execute multi-step plans, not just generate clever one-off responses. For those exploring engineering-grade AI workflows, this distinction between generating improvements and actually benefiting from them is critical.
The Bigger Picture
This work arrives as enterprises move from experimenting with AI to deploying it in production. The gap between "this model is smart" and "this model reliably improves our workflow" is wider than many assume. Self-evolving agents sound futuristic, but the research suggests the real unlock isn't autonomous self-improvement—it's designing systems where mid-tier models can consistently leverage structured guidance.
The code is open-source, meaning developers can test these patterns in their own agent systems. The practical takeaway: invest in harness design and instruction-following capability before chasing self-evolution features.
