AutoScientists puts a decentralized AI lab on autopilot

The biggest bottleneck in computational research usually isn't a lack of ideas. It's the sheer amount of compute wasted when researchers run down dead ends that someone else already tried.

AI agents trying to do science have suffered from the exact same problem. Most existing systems either follow a rigid single trajectory or rely on a centralized planner with fixed instructions, making it nearly impossible to sustain parallel exploration or remember past mistakes.

Researchers introduced AutoScientists, a decentralized multi-agent system built specifically for long-running computational scientific experimentation.

Ditching the central micromanager

Instead of trusting one master agent to dictate every step, AutoScientists lets agents read a shared experimental state and self-organize into teams around the most promising hypotheses.

Crucially, they don't just burn GPUs on every wild hunch. The agents critique each other's proposals before executing any code, filtering out flawed setups before compute is spent.

They also log and broadcast their failures. When an experimental direction hits a brick wall, the entire system learns about it, preventing other agents from running redundant tests on dead hypotheses.

Crushing benchmarks on equal budgets

Under identical compute budgets, the decentralized approach comfortably beat previous single-agent setups across multiple complex domains.

On BioML-Bench, which spans 24 computational biology tasks across drug discovery, protein engineering, single-cell omics, and biomedical imaging, AutoScientists scored an average leaderboard percentile of 74.4%. That is an 8.33% improvement over the previous best AI agent.

When optimizing GPT training pipelines, AutoScientists hit the target validation bits-per-byte 1.9x faster than Autoresearch. Even more telling: starting from an already optimized baseline where single-agent approaches stalled out completely, AutoScientists found 7 accepted improvements where the baseline found zero.

In ProteinGym fitness prediction, it engineered a novel method for ACE2-Spike binding that beat the state of the art by 12.5% in Spearman correlation. Applied out of the box across all 217 ProteinGym assays, the exact same method improved over the prior SOTA by 6.5%.

Why it matters

We think AutoScientists signals a major shift in how AI labs will automate discovery. Instead of building bigger, more expensive single agents, the smarter path is turning agents into a self-checking scientific community.

The real breakthrough here is failure sharing. In long scientific runs, knowing what not to do is half the battle. By forcing agents to peer-review proposals and track dead ends, AutoScientists turns wasted trial-and-error into collective intelligence.

For computational biologists and ML engineers, this model offers a tangible glimpse into automated R&D. It's not yet ready for wet-lab physical pipetting, but for pure in-silico discovery, letting a swarm of specialized agents argue over code and share notes looks far more effective than trusting a single prompt to do it all.