Science agents can run the workflow. Proving the result is the hard part

Tools work. Checking is the gap

A benchmark built for scientific discovery agents tests whether agents can recover governing laws, not only run investigations. SciGame uses structured cards pairing original environments with dimensionally consistent counterfactuals, in a Historical collection of canonical laws and a live Temporal collection curated from papers published after model knowledge cutoffs. Familiar priors hurt agent performance once the governing relation is altered, and strict success on Historical counterfactuals was 0 percent for all five Codex models tested. Active investigation alone, the benchmark suggests, does not recover the law when the world behind the data changes. m:…3395

But a preprint arguing for a shift in how science agents are judged warns that running tools is not the same as doing science. Its authors name environmental verifiability as the central systems constraint on agent autonomy: evidence must be timely, attributable, and affordable, or the agent cannot tell whether a new observation supports or challenges its hypothesis. Workflow automation alone, they argue, does not constitute scientific discovery. The distinction matters for anyone selling an agent to a laboratory. A system that executes a protocol flawlessly but cannot say what would count as evidence against its own conclusion is a workflow runner, not a research partner, and its failures will look like confident, well-formatted errors. m:…8903

Discovery games expose the proof gap

The failure mode is instructive. Approximately 48 percent of one model's failures involve near-perfect explanations without a symbolic equation: the agent can describe the result well while failing to identify the underlying equation. Adaptive experimentation helps the strongest models but harms weaker ones under limited budgets. That combination matters for deployment: explanation quality is not proof, and giving an agent more experiments helps only if the agent can use them within budget. m:…3395

Labs that adapt, evidence that must be checked

A self-evolving laboratory system shows the execution side of that division. GPT Astra autonomously generates and 3D-prints labware, including beaker holders and storage stands, based on physical feedback and camera observations, adapting its physical and digital infrastructure to experimental needs. The system was validated through a photoresist polymerization campaign using a 405 nanometre laser, with ATR-FTIR showing reduced acrylate-band absorbance. Adaptation is therefore not only software: the lab itself changes shape to fit the workflow, which makes the verification question more urgent, since a changed setup must still produce checkable evidence. m:…7343

Read together, these studies describe a maturing division of labor. Agents can investigate counterfactual worlds and adapt physical labware, but the evidence behind each step must be timely, attributable and affordable enough for a human expert to check without redoing the work. Until that check is cheap, the sensible deployment keeps experts in the loop at exactly the points where verification is hardest: recovering a symbolic law, recording a result, and adapting a physical setup. The open question is not whether agents can run scientific workflows. It is whether the proof behind each step can be checked. The perspective paper behind that gap proposes a three-stage framework of reasoning, execution, and world-interaction, noting that current agents excel at digital workflow execution while lacking the capacity to integrate physical experimental evidence into iterative hypothesis revision. A laboratory that prints its own holders is therefore only as trustworthy as the check on what those holders enabled. m:…8903 m:…3395