Equipping Agentic AI for Scientific Research
Mirror was founded to unlock new scientific capabilities made possible by the confluence of breakthroughs in scientific simulation and the emergence of sophisticated reasoning systems. As these technologies develop, their potential to disrupt how scientific research is performed is becoming increasingly clear — even inevitable.
As we’ve experimented with existing AI tools and agents at Mirror, however, we’ve found them to be essentially unequipped to support real scientific work. Despite demonstrating strong potential in niche research applications, they’re often costly, brittle, unreliable, and difficult to implement and integrate into scaled workflows.
There’s a pressing need to lower barriers around using state-of-the-art tools so that scientists can focus and act on their ideas effectively, efficiently, and scalably. Computational science should enjoy the same level of fluid, general-purpose autonomy that AI has brought to software engineering. Executing a novel workflow should be a matter of outlining scientific directives in a few minutes, not spending hours to days battling the machinery behind implementation.
Introducing Axon
To facilitate this new approach to science, we’ve created Axon, an interactive AI tool that plans, executes, and analyzes computational chemistry workflows via natural language.
At its core, Axon is a system that gives reasoning agents access to a diverse collection of computational scientific tools as well as the requisite compute infrastructure to automate and deploy them at industrial scale. Axon benefits from a first-of-its-kind knowledge acquisition pipeline and guidance system that enables it to quickly onboard new tools, learn from previous mistakes, and adapt to new integrations or unfamiliar contexts, without ever needing to train on proprietary data or retain sensitive information. Simultaneously, work done with Axon is inspectable down to decision traces and individual calculations. Everything is neatly packaged for seamless deployment, but we’ve taken care that nothing is a black box.
Axon is currently designed for drug discovery, a field with significant computational friction despite widespread adoption of AI- and simulation-based tooling. Our early experiments suggest that Axon can significantly reduce friction in pharmaceutical workflows through its fluency in wielding powerful computational tools that scientists already use and trust, now with the harnessing and infrastructure to deploy these tools at scale.
What can Axon do?
We tested Axon’s ability to execute a variety of tasks spanning the drug discovery pipeline (docking and virtual screening, protein folding, molecular dynamics, ADMET modeling, molecule generation, etc.) to begin scoping its function and utility. Importantly, all of these tasks were prompted using only natural language, and for each, Axon invoked industry-standard third-party tools and automatically configured compute resources as needed.
We initially tested a variety of platforms (Biomni, ChatGPT, Kosmos) on the same tasks, but found that Claude Code was the only alternative to admit a meaningful comparison. Nevertheless, even against Claude Code, we found that Axon executes computational drug discovery tasks better, faster, and at lower cost.
The Details
These tasks were designed as regression tests to track Axon’s performance throughout development. As such, for speed and clarity, tasks were designed to be short and representative (e.g., 10 instead of 10k steps of MD), each covering a single stage of a computational workflow (e.g., “generate a Packmol input file” or “dock these ligands against my target protein”). The intention is for these tasks to test Axon’s executional abilities, not to evaluate the underlying scientific tool. Accordingly, these tests do not yet probe reliability on extended, multi-hour work horizons. We evaluated results after a ten minute cutoff, although smooth executions were expected to complete well before.
Following each task, we gave an evaluator (Claude Opus 4.5) access to a transcript of the agent’s run and any additional output files. The evaluator was directed to grade Axon (or other agent) on execution quality (efficiency, justified exploration), reasoning quality (clarity, planning), and task completion. We then considered each of these scores individually, a performance metric as the average of the three, time required for each run, and cost as the total monetary cost of each run including API-based LLM calls and sandboxed GPU resources.
We set a very high standard for what is considered a successful run: a task completion score of 9 or higher out of 10. Since the tasks were small units of work already, whether a task ran to completion (regardless of scientific correctness of the outcome) provides clear signal of whether a workflow could be continued following the indicated task. Runs that took longer than 10 minutes were terminated and accordingly received poor task completion scores, as long runtimes were generally indicative of agents getting stuck in environmental configuration spirals. In practice, these runs would likely never complete successfully or would significantly hinder computational workflows.
The Results
Axon successfully completed 26 out of the 27 tasks. Claude Code with Opus 4.5 was the best performing of the non-Axon agents we tested on the same tasks. Even so, Claude Code only completed two-thirds of the tasks (18/27) and failed to produce any outcome for many of the rest, scientifically correct or otherwise. This means that Axon was able to complete 44% more tasks than Claude Code, demonstrating significantly higher efficiency in time and resource usage.
If you sum up total spend across the regression tests, it appears that Axon was more expensive than Claude Code. However, because of its poor completion ratio, Claude Code actually wasted notably more time and compute on dead ends—16⨉ more time and 17⨉ more spend than Axon used on non-completions. This means that Axon used more time and compute overall, but effectively applied these resources towards achieving better performance and real results. Only successfully completed tasks can drive research forward; accounting for these completions, Axon was 3⨉ faster and 30% cheaper on average to completed tasks. Intuitively, this means triple the throughput of useful results on your computational chemistry workflows at lower cost compared to other state-of-the-art agents.
Axon also scores better on average when examining overall score, which also takes reasoning and execution processes into account. Let’s look into a few specific task examples:
Docking calculation: “Using AutoDock-GPU, perform a basic molecular docking of biotin to streptavidin. Use the receptor grid maps at “receptor.maps.fld” and the ligand at “ligand.pdbqt”. Run 5 independent docking runs. Save the DLG and XML output files locally.”
Axon achieved a performance score of 8.5/10 and cost $0.80, while Claude Code achieved only 4/10 performance at a cost of $3.51. Based on the run transcripts, both agents understood the task and found tool documentation to understand the CLI interface. However, Axon proceeded to complete the task with minimal friction, quickly resolving a minor error with directory structure. Meanwhile, Claude Code encountered significant kernel compilation issues and did not properly diagnose them. Instead, Claude Code tried several minor variations of approaches that consistently failed and made no meaningful progress toward the stated goal.Molecular descriptors: “Using Rowan, calculate molecular descriptors for the molecule with SMILES “CC(=O)Oc1ccccc1C(=O)O”. Submit the calculation, wait for completion, and return the computed descriptors including molecular weight, logP, and topological polar surface area.”
Axon and Claude Code both successfully in completed this task and demonstrated strong overall performance (9.2/10 and 8.3/10, respectively). Claude Code received minor deductions for unclear presentation of its final answer, lack of diagnosis for an error it encountered, and inefficient troubleshooting to get the API call working correctly. Ultimately, these errors translated to higher costs and longer runtime ($0.52, 171 s) compared to Axon ($0.42, 78 s).LAMMPS simulation: “Using LAMMPS, run a short (ten step) molecular dynamics simulation of Lennard-Jones particles. Use the input script at “in.lj_melt”. Verify the simulation completes successfully and extract thermodynamic data (temperature, pressure, energy) from the log file.”
This task was the only one Axon failed (6/10 task completion, $0.79 spent, 88 s runtime). Despite reasonable understanding and planning, Axon set up LAMMPS successfully but failed to modify the run length, and so ran 100 steps rather than the instructed 10 and did not extract the thermodynamic data for the user. Claude Code comprehensively completed the task with reasonable spend (10/10 completion, $0.58 cost), but was inefficient in installing packages and ironing out library dependencies (139 s runtime).
Over many more tasks, Claude Code was primarily limited by infrastructural blocks, and failing to navigate around them. In some cases, those issues were not Claude Code’s fault—it lacked access to compute resources and could not run certain installations despite using the most relaxed permissions settings available. Some of these restrictions are in place to protect the users’ data and compute environments and cannot be overridden in API-based usage of Claude Code.
Axon was built precisely to provide the infrastructure and harnessing to safely and efficiently circumvent these challenges; to enable agents to navigate and execute complex computational tools at scale. Our testing so far demonstrates that Axon can identify how and when to use its tools accomplish the task at hand, manipulating inputs and resources to achieve that goal. In most cases where Axon displayed limited performance, it was attributable to a limited understanding of tools and their failure modes — a limitation which is easily corrected with improved tools and the accumulation of experience and knowledge. We’ve built an an internal system for automatically synthesizing and organizing this experience, which we'll dive into in a future post.
Future Axon
Of course, there is always room for improvement. And importantly, these are early results for Axon, the first we were excited to share. There are many more tools used in drug discovery that weren’t part of our regression tests because open-source options weren’t available, or because the integrations were more complex and we simply haven’t gotten to them. We are working quickly to add more tools to improve the breadth of skills available to Axon. Nevertheless, looking towards the utility of Axon for real scientific work, the tasks we showed here consist of small steps that can be strung together to make complete workflows in drug discovery. While we give Axon one step at a time here, it is also well capable of executing extended task lists and of generating its own. We’re actively working on benchmarking many-step execution under “higher-order” objectives, so stay tuned for more!
To ensure that we are harnessing state-of-the-art tools in quickly evolving fields, we’ve designed Axon to be a self-improving platform. We’ve demonstrated internally that, with relatively straightforward tools, Axon can find documentation and independently run the end-to-end process of installing, on-boarding, and successfully deploying new tools.
Additionally, Axon retains its own knowledge base that it adds to as it explores new domains, learns new tools, and orchestrates different workflows. This retention means that important context it learns as it performs more—and more diverse—tasks is always available. Importantly, Axon can learn from outcomes alone: observing success and failure without retaining or retraining on the actual content, so propriety data remains entirely confidential in enterprise deployments. All this means that Axon continually improves with use, and in instances where Axon displays significant knowledge or performance gaps, users can manually add to the knowledge base or apply reinforcement learning pipelines to help Axon tune new skills.
We’ve tailored Axon to support the requirements of drug discovery because the maturity and wide adoption of complex computational tools in the field marks a near-term path to value for streamlining and automation. As we continue to improve Axon, there is clear extensibility to other domains, and we can’t wait to put these transformative capabilities in the hands of more scientists.
We're super excited about this tech, and have way more to say about it than fits in a blog post. If you have questions or comments or just want to talk with us, shoot me a message at sabrina@mirrorphysics.com.