Lab Automation, Now in Plain English

by

Author: Denis Avetisyan

Researchers have developed an AI agent that translates natural language instructions into executable laboratory protocols, streamlining experiment design and execution.

A system facilitates iterative refinement of experimental protocols, where an intelligent agent-informed by laboratory context-translates structured text definitions → into visual representations, allowing for bidirectional modification by researchers and automated error correction, embodying a cycle of assisted evolution rather than static design.
A system facilitates iterative refinement of experimental protocols, where an intelligent agent-informed by laboratory context-translates structured text definitions → into visual representations, allowing for bidirectional modification by researchers and automated error correction, embodying a cycle of assisted evolution rather than static design.

This work presents an AI-powered agent integrated with laboratory orchestration software, enabling scientists to create, monitor, and analyze experiments using simple text prompts.

Despite the potential of automated laboratories to accelerate scientific discovery, realizing this vision requires significant effort in coding and configuring complex systems-a barrier for many researchers. This challenge is addressed in ‘From Prompts to Protocols: An AI Agent for Laboratory Automation’, which introduces an AI agent architecture integrated with laboratory orchestration software to translate natural language instructions into executable experimental protocols. The system achieves a 97% first-attempt success rate in protocol generation and dramatically reduces required user interactions, effectively bridging the gap between high-level experimental goals and automated execution. Could this approach unlock a new era of accessible and efficient scientific experimentation, empowering researchers to focus on hypothesis generation rather than laborious automation tasks?

The Inevitable Bottleneck: Reframing Experimental Design

Scientific advancement is frequently constrained not by a lack of ideas, but by the laborious process of translating those ideas into rigorously controlled experiments. Traditional experimental design demands substantial manual effort – from meticulously planning each step of a protocol to painstakingly executing it, and finally, to carefully recording and analyzing the resulting data. This process isn’t merely time-consuming; it requires a high degree of specialized expertise to ensure accuracy and reliability. The sheer volume of manual intervention creates a significant bottleneck, slowing the pace of discovery and limiting the ability of researchers to rapidly iterate on hypotheses. Consequently, promising avenues of research may be delayed or even abandoned due to the practical difficulties inherent in setting up and running complex experiments, highlighting the need for innovative approaches to streamline this critical stage of the scientific method.

The painstaking process of defining and executing intricate experimental protocols represents a significant impediment to the pace of scientific advancement. Researchers often dedicate substantial time to meticulously outlining each step, from reagent preparation to data acquisition, a task susceptible to human error at any stage. These errors, however minor, can invalidate results, necessitating repetition and delaying crucial insights. The manual nature of protocol execution further restricts the speed of iteration; responding to unexpected findings or adapting to changing conditions requires significant intervention and reconfiguration, effectively creating a bottleneck between observation and further investigation. Consequently, the potential for rapid discovery is diminished, as the emphasis shifts from exploration to error correction and procedural refinement.

Modern research increasingly demands experimental frameworks capable of dynamic adjustment and iterative refinement, a need driving the development of automated and adaptable protocols. Traditional methods, often reliant on static, pre-defined procedures, struggle to accommodate the complexities of contemporary scientific inquiry and the sheer volume of generated data. The capacity to rapidly modify experimental parameters in response to real-time results – or even to autonomously explore alternative pathways – promises to dramatically accelerate discovery. Such systems not only reduce the potential for human error but also enable researchers to tackle problems previously intractable due to logistical or temporal constraints, fostering a more agile and efficient scientific process. This shift towards programmable experimentation represents a fundamental change, moving beyond simply executing protocols to actively optimizing them for maximum insight.

Existing experimental workflows often operate as rigid sequences, proving inadequate when faced with the inherent unpredictability of scientific investigation. A study of automated microfluidic systems revealed that deviations from anticipated outcomes frequently necessitate complete experimental restarts, consuming valuable resources and time. This inflexibility stems from protocols designed for pre-defined scenarios, lacking the capacity to dynamically adjust parameters or branching logic in response to real-time data. Consequently, unexpected results-which often represent the most exciting discoveries-are frequently overlooked or dismissed due to the inability of current methods to adapt and explore alternative pathways. The development of protocols capable of self-correction and adaptive learning is therefore paramount to accelerating the pace of scientific advancement, enabling researchers to capitalize on serendipitous findings and navigate the complexities of biological systems with greater efficiency.

The EOS AI agent empowers scientists to design, execute, and analyze experiments using both natural language and an interactive visual graph editor.
The EOS AI agent empowers scientists to design, execute, and analyze experiments using both natural language and an interactive visual graph editor.

Bridging the Gap: An Intelligent Agent for Protocol Generation

The AI Agent employs Large Language Models (LLMs) to interpret experimental objectives communicated through natural language input. This capability bypasses the need for researchers to manually translate goals into machine-readable code or predefined templates. The LLM processes user-provided descriptions of the desired experiment – including specifications for reagents, conditions, and expected outcomes – and extracts the essential parameters necessary for protocol generation. This natural language understanding allows for increased flexibility and accessibility, enabling researchers to define experiments in a manner that closely mirrors their thought process, which is then converted into executable instructions for automated laboratory equipment.

The AI Agent functions as an intermediary between high-level experimental goals and the physical execution of laboratory procedures. It achieves this through direct integration with existing laboratory orchestration systems – encompassing robotic liquid handlers, temperature controllers, and data acquisition hardware. Upon receiving a protocol defined in YAML, the agent parses the instructions and translates them into a series of actionable commands compatible with the connected instrumentation. This translation process includes mapping parameter values, sequencing operations, and managing data flow, effectively automating the experimental workflow without requiring manual reprogramming of the underlying hardware.

Protocol definition within the AI Agent is managed using YAML (YAML Ain’t Markup Language), a human-readable data serialization format. YAML’s structure, utilizing key-value pairs and nested lists, facilitates clear organization and modification of experimental parameters. This contrasts with more complex formats like JSON or XML, offering improved readability and ease of editing for researchers. The use of YAML allows for straightforward version control, collaborative editing, and automated parsing by the agent, enabling rapid protocol iteration and adaptation without requiring specialized programming skills. The format supports complex data types, including strings, numbers, booleans, and lists, providing the flexibility needed to define a wide range of experimental procedures.

The AI Agent employs Bayesian Optimization as a core component for automated protocol refinement. Across 65 experimental trials, conducted on four distinct problem sets, this approach yielded a 97% first-attempt success rate. Bayesian Optimization facilitates the efficient exploration of the protocol parameter space by constructing a probabilistic model to predict performance based on prior evaluations. This model is then used to intelligently select the next protocol configuration to test, balancing exploration of uncertain areas with exploitation of promising configurations, ultimately minimizing the number of iterations required to achieve optimal performance.

This demonstration showcases a conversational interface allowing users to load experimental protocols, laboratory code, and initiate closed-loop optimization campaigns with the EOS AI agent.
This demonstration showcases a conversational interface allowing users to load experimental protocols, laboratory code, and initiate closed-loop optimization campaigns with the EOS AI agent.

Visualizing the Process: Empowering Scientists Through Transparency

The Visual Graph Editor utilizes a node-based diagrammatic interface for protocol construction and modification. This interface represents experimental procedures as a network of interconnected nodes, each representing a specific operation or data transformation. Users manipulate these nodes and their connections visually, defining the flow of data and control within the protocol. This approach facilitates rapid prototyping and iterative refinement of experimental workflows by providing a clear, graphical representation of complex procedures, eliminating the need for extensive coding or scripting. The editor supports drag-and-drop functionality, customizable node properties, and real-time validation of protocol connectivity, enhancing usability and reducing the potential for errors.

Real-time protocol monitoring within the system provides experiment progress tracking through continuous data acquisition and analysis. This functionality displays key experimental parameters and metrics as they change over time, allowing scientists to observe the execution of each protocol step. The system then utilizes this data to provide immediate feedback on protocol performance, identifying potential deviations from expected results. This enables dynamic adjustments to experimental parameters – such as reagent concentrations, incubation times, or temperature settings – during runtime, optimizing the experiment and minimizing the need for re-runs. The collected data is also logged for post-experiment analysis and reproducibility.

The system incorporates data analysis capabilities that facilitate the extraction of statistically relevant insights from experimental datasets. These analyses include, but are not limited to, quantitative measurements of key performance indicators, identification of correlations between variables, and the generation of visualizations such as charts and graphs. Results are presented in a standardized format allowing for direct comparison between experimental runs and enabling scientists to refine protocols and parameters for subsequent iterations, ultimately accelerating the pace of scientific discovery. The system supports multiple data export formats for integration with external analysis tools and long-term data archiving.

Successful integration of the system onto the PurPOSE Platform demonstrates its adaptability to existing research infrastructures. The PurPOSE Platform, a distributed research environment focused on synthetic biology, required a modular and scalable solution for protocol design and execution. The system’s implementation facilitated the automation of complex biological workflows within PurPOSE, validating its compatibility with diverse hardware and software components. This deployment involved adapting the system to handle PurPOSE’s specific data formats and communication protocols, confirming its flexibility beyond the initial design specifications and establishing a functional proof-of-concept for broader adoption in similar research settings.

The EOS AI agent seamlessly integrates with the EOS user interface, enabling both automated protocol creation within the protocol editor and AI-assisted analysis of optimization campaign progress.
The EOS AI agent seamlessly integrates with the EOS user interface, enabling both automated protocol creation within the protocol editor and AI-assisted analysis of optimization campaign progress.

From Optimization to Extraction: A Glimpse of Accelerated Discovery

The AI agent demonstrated a remarkable capacity for optimization through a simulated color mixing challenge. Across 35 trials, the system successfully formulated the correct color mixture on 33 occasions-a 94% first-attempt success rate. This achievement wasn’t simply about random chance; the agent learned to navigate the complex relationships between input color components to achieve precise target colors. This proficiency indicates a powerful ability to solve problems requiring nuanced control and iterative refinement, suggesting potential applications in fields demanding precise compositional control, such as materials science and chemical synthesis.

The AI agent successfully performed automated liquid-liquid extraction, a complex separation technique crucial in chemistry and materials science. This demonstration involved the precise manipulation of fluids to selectively transfer target compounds based on solubility, showcasing the system’s ability to control experimental parameters and achieve desired separations without human intervention. By autonomously managing variables like solvent ratios, mixing speeds, and phase separation, the agent proved capable of executing a sophisticated process typically requiring significant manual skill and optimization. This achievement underscores the potential for automated experimentation to accelerate the discovery of new materials and chemical processes, offering a pathway to faster, more efficient research cycles.

The AI agent’s capabilities extended to solubility screening, a crucial process in early-stage drug discovery and materials science. This involved the system autonomously designing and executing a comprehensive assay to determine the solubility of various compounds under differing conditions. Rather than relying on pre-defined protocols, the agent formulated experimental plans – including reagent selection, concentration gradients, and measurement timings – to maximize information gain with each trial. Validation confirmed the system’s ability to not only generate viable screening procedures, but also to accurately interpret results and refine subsequent experiments, demonstrating a closed-loop optimization of the entire assay process and highlighting its potential to dramatically accelerate materials and pharmaceutical development.

The demonstrated capabilities extend beyond isolated tasks, suggesting a paradigm shift in the pace of scientific investigation. By automating experimental workflows – from optimization and extraction to solubility screening – the system drastically reduces the cognitive load and time commitment traditionally required for research. Analysis indicates a reduction in interaction complexity, ranging from nine to twenty-seven times less than conventional manual authoring of experiments. This efficiency gain isn’t merely incremental; it unlocks the potential for researchers to explore a far wider experimental space, test hypotheses more rapidly, and ultimately, accelerate the rate of discovery across diverse scientific disciplines. The approach promises to reshape how experiments are designed, executed, and analyzed, fostering innovation and pushing the boundaries of knowledge.

The EOS AI agent successfully implements a color mixing protocol to achieve desired hues.
The EOS AI agent successfully implements a color mixing protocol to achieve desired hues.

The pursuit of automated laboratory workflows, as detailed in the paper, inherently acknowledges the transient nature of any complex system. The agent’s ability to translate natural language into executable protocols represents a momentary stabilization against entropy – a caching of desired functionality against the inevitable decay of manual processes. As Ada Lovelace observed, “The Analytical Engine has no pretensions whatever to originate anything. It can do whatever we know how to order it to perform.” This echoes the AI agent’s reliance on defined protocols; it excels at execution, but the ingenuity still resides in the scientist’s initial design. The agent simply accelerates the process, offering a temporary illusion of control over the relentless march of time and complexity within the laboratory environment.

What’s Next?

The integration of large language models with laboratory automation, as demonstrated, isn’t a resolution, but a reshuffling of dependencies. Each iteration of this agentic loop-from prompt to protocol-records a commitment in the annals of scientific tooling, and every version represents a chapter in a perpetually drafted manuscript. The current system addresses the friction of interface, yet the fundamental constraints of experimental design remain. The true tax on ambition isn’t the time spent coding, but the inevitable limitations of the models themselves-their susceptibility to hallucination, their opaque reasoning, and their dependence on the quality of training data.

Future development will likely center not on achieving perfect automation-an asymptotic ideal-but on building systems that gracefully degrade. The challenge lies in designing agents capable of recognizing their own limitations, of requesting human intervention with appropriate context, and of learning from those corrections. An agent that flags uncertainty is, paradoxically, more valuable than one that confidently delivers error.

The field will inevitably confront the question of reproducibility. If an experiment is orchestrated by a language model, the ‘method’ section becomes a complex record of prompts, parameters, and model versions-a provenance that demands careful curation. Every commit is a record, and maintaining that history is the price of admission to a future where experiments aren’t merely executed, but understood-not just by the machine, but by the scientist who initiated the query.

Original article: https://arxiv.org/pdf/2605.16552.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

See also:

2026-05-19 11:34