ADAS: a meta-agent that invents new AI agent architectures
In one sentence University of British Columbia publishes ADAS (Automated Design of Agentic Systems): a meta-agent that searches for new agent architectures by writing and evaluating Python code. Discovers novel patterns (dynamic critic, step-back abstraction) that outperform human-designed agents. First system automating agent architecture research.
Usually AI agent architectures are designed by human researchers: someone studies the problem, proposes an idea ("let a second agent critique the answer"), implements it and tests it. It is a slow process.
ADAS does this automatically. It is a meta-agent, meaning an agent whose goal is to find better agents. It works like this: given a benchmark task, ADAS writes Python code to describe a new agent architecture, runs it on the benchmark, measures performance, notes the results, and then generates a new architecture based on what it has learned.
During this autonomous search process, ADAS discovered patterns that human researchers had not yet codified: for example a "dynamic critic" that activates only when the main agent expresses uncertainty, or a "step-back abstraction" technique that makes the agent take a step back on a problem before solving it in detail.
It is the first system that automates research on agent architectures themselves: meta-AI designing AI.
Companies
University of British Columbia
Tools
—
Tags
Sources