收录解读
Automated Design of Agentic Systems reframes a large part of agent engineering as a search problem rather than a hand-designed craft. Instead of manually inventing prompts, tool-use patterns, or workflows, it asks whether a meta-agent can program and iteratively discover stronger agent systems directly in code.
The main novelty is elevating automatic agent design into an explicit research area, ADAS, and grounding that idea with Meta Agent Search. The paper argues that if agent systems are code, then the search space includes prompts, control flow, tool use, and system composition all at once, making learned design a plausible replacement for hand-built architectures.
This matters for the repository because it changes how one frames progress in agent systems. It is not merely another workflow-search paper; it is a route-level paper that says the design of agentic systems itself should become an optimization target. That has strong spillover into autonomous science, workflow generation, and self-evolving systems.
It is not ranked higher because the line is still early and expensive, and later work is still testing its practical limits. But the framing is important enough that it deserves formal collection as a durable reference.