Multi-agent systems get an optimizer, plus a reality check from real deployments

Topology search beats heuristics

One preprint introduces CATS, contrastive attribution topology search, for multi-agent systems. The idea is to decide which agents belong in the system and how they connect based on measured attribution rather than gut feel. On BigCodeBench, CATS improves mean accuracy by 5.3 percentage points over score-only heuristic search while cutting deployment token usage by 54%. In HotpotQA stress tests, the same method successfully excludes adversarially prompted agents: it identifies and drops the saboteur instead of letting it poison the ensemble. The method also functions as a filter: when an agent contributes nothing, or actively sabotages, attribution-based search removes it rather than retuning the prompts around it. m:…1951

For founders, the interesting number is the second one. A 54% cut in deployment tokens is a direct margin improvement on every agent call, and it arrives alongside better accuracy rather than trading against it. The implication, stated here as reasoning rather than reported fact: agent-system performance is increasingly a search problem, not a prompt-engineering problem, and the economics of that search now favor whoever optimizes topology first. m:…1951

Bayesian evolution over LLM guesses

A second preprint takes a different optimizer to the same problem. Bayesian evolutionary optimization, or BEO, gives multi-agent topology discovery an explicit selection criterion that outperforms LLM-based candidate selection. The evolved architectures achieve the highest overall performance on BigCodeBench among the agent systems the authors evaluated. The same framework's optimized preprocessing improves TabPFN performance on temporal tabular datasets while remaining competitive on IID datasets, which suggests the tuning generalizes beyond code tasks. That a hand-tuned team can be beaten by an explicit search criterion should be read as good news: team design is becoming reproducible, which means it is becoming shippable. m:…1739

Read together with CATS, a pattern emerges: the field is moving from asking a language model to pick the team to optimizing the team with a proper objective. For anyone building agent infrastructure, the question shifts from which agents to prompt to which search procedure to run. Both papers arrive the same week, both beat heuristic baselines, and both treat the ensemble itself as the thing to engineer. m:…1739

What real deployments actually use

The third paper grounds both of the above. It is a systematic mapping study of multi-agent LLMs across the cloud-edge-IoT continuum, identifying five recurring architectural families, including specialized agent teams and federated model adaptation, and six deployment patterns ranging from local-only execution to policy-gated control loops. The reality check: physical hardware deployments are dominated by open-source models with fewer than 8 billion parameters. The frontier demos run big models; the deployed systems run small, open ones. m:…7788

That gap between demo and deployment is the founder's opportunity and the founder's constraint. Topology optimizers like CATS and BEO are benchmarked on capable models, but the mapping study says the hardware reality is sub-8B open weights. The winning product, reasoning from both facts, is the one whose topology search works under the token and memory budgets that real deployments actually have. Founders who benchmark only on frontier models are optimizing for a deployment environment most customers do not have; the mapping study is a reminder to test where the workloads actually run. m:…7788