The 2023–2026 shift from chatbots to closed-loop lab agents — what's been demonstrated, what's still human, and what it means for your bench.
In narrow, well-bounded problems, yes — AI “agents” have already planned and run real experiments and produced validated results. In 2023 an LLM agent optimised real chemical reactions; in 2025 a team of LLM agents designed SARS-CoV-2 nanobodies that were then confirmed at the bench. But every success shares one pattern: a human sets the goal and reviews the output, and the wins come where the objective and the success signal are crisp. Deciding which experiment is worth running is still human.
A chatbot answers on a screen. A lab agent wraps a language model with three extra powers: tools (literature search, code, simulation), memory, and an interface to real instruments. That turns “describe the experiment” into plan it, run it, read the result, and decide the next step — a closed loop. The jump from 2023 to 2026 is mostly this: from one model answering, to teams of role-specific agents (a “PI” agent directing “scientist” agents) doing multi-step work.
The 2025 Virtual Lab result is the milestone: a team of LLM agents, with light human steering, produced experimentally confirmed biology — including nanobodies with improved binding to recent SARS-CoV-2 variants [4] — not just a plausible plan. (A caution from the same era: A-Lab's novelty claims were later questioned and corrected [3].)
Notice what all three share — a clearly stated goal, defined tools, and an unambiguous success signal (did the reaction work? did the nanobody bind?). That's exactly where agents shine. The unsolved part is upstream: choosing a novel target, inventing a new modality, and diagnosing an ambiguous failure that fits no expected pattern.
The field is converging on human-in-the-loop autonomy — AI does the data-heavy planning and execution while an experienced scientist keeps the real-time decisions [5]. The destination most labs are steering toward isn't a lights-out megalab; it's a self-driving bench you own, where you describe the protocol and the robot runs it, vision-verified at each step. See how that works →
