AGTBRIEF
OptiMAS automates the trial and error of building multi agent LLM systems, reporting competitive or better results than hand crafted setups on four benchmarks.
An optimizer reads the transcripts and task scores from a team of AI agents, then proposes edits to the team. A dual-track memory is what the authors credit for keeping that loop improving over many rounds.
That is the core of OptiMAS, a task-agnostic tuner for multi-agent LLM (large language model) systems, in which several AI agents collaborate on a task by passing messages and tool calls. The paper, posted to arXiv this month, frames OptiMAS as continuous, data-driven optimization on top of a ReAct-based infrastructure. ReAct is a common agent pattern in which the model alternates between natural-language reasoning and concrete tool actions.
The authors describe a trade-off in earlier search-based methods. Expanding the search space made evolutionary runs unstable, while discrete branch-and-discard search kept useful insights trapped in separate lineages. OptiMAS sidesteps that by treating textual interaction traces and task feedback as the loss signal for end-to-end evolution, with a dual-track memory the authors say preserves useful prior behavior across long optimization horizons.
The team evaluated OptiMAS on four heterogeneous agentic benchmarks (multi-step task suites scored by an automated grader) across three LLM backbones of varying scale and access, and reports results "competitive or superior" to hand-crafted, domain-specific systems and earlier evolutionary approaches (full paper).
The limits matter. Four benchmarks and three backbones is a thin evidence base. The numbers are the authors' own, with no independent reproduction visible. "Competitive or superior" is the paper's framing, not an outside verdict. Wire copy will probably say "AI improves AI." The actual story is a closed-loop transcript-plus-score tuner for agent teams: useful, scoped, and unsettled.