Since the beginning of 2026, the rise of OpenClaw has marked a significant shift in large language model (LLM) capabilities—from merely chatting with humans to operating in the physical world. It also signals the democratization of AI agent technology, which was previously accessible only to a handful of companies or individuals. AI agents are now empowering industries across the board, and drug discovery is no exception.
AIDD (AI for Drug Discovery) represents the intersection of AI and pharmaceutical research. From 2019 to the present, drug discovery researchers have witnessed a rapid evolution in how they interact with AIDD models—from command-line interfaces (CLI) to web applications, and now to intelligent agents. What remains constant is the pursuit of efficiency, ease of use, and accessibility.
Today, we can boldly declare that a new era of AIDD agents has arrived. No longer do researchers need to type complex commands in Terminal or click through SaaS interfaces with a mouse. With natural language alone, scientists can now invoke AIDD models for virtual screening or drug design.
But is the AIDD agent revolution merely about changing the user interface? We believe it goes far beyond that. In this article, we present six predictions for this new era.
Prediction 1: AIDD Base Model Usage Will Increase 10-Fold or More
Agents that operate through natural language dramatically lower the barrier to using AIDD tools. When biologists can analyze protein structures using AlphaFold through simple natural language commands, and when chemists can generate molecular libraries by verbally describing their needs—AIDD transforms from a "specialized tool" into "infrastructure." Just as Excel enabled non-programmers to perform data analysis, AIDD agents will give every scientist their own computational chemistry assistant.
We predict that a large number of biologists and chemists without AIDD/CADD backgrounds will begin using AIDD base models for their research. The number of users willing to deploy models locally or call cloud-based APIs will increase significantly, and usage will become more frequent.
Prediction 2: AIDD Agents Will Autonomously Plan Workflows to Address Long-Tail Drug Discovery Scenarios
With their reasoning and autonomous planning capabilities, agents will break the rigid computational workflows of traditional SaaS platforms. We will see various AIDD models being combined in unexpected ways to solve highly specific drug discovery problems—such as developing dual-target or triple-target agonists, or creating cross-species antibodies. These scenarios were previously overlooked by platform companies due to their rarity.
Traditional SaaS business models focus on "covering the greatest common denominator"—building only the features that 80% of users need frequently. But in the agent era, a graduate student can describe their unique needs in natural language, and the agent will automatically combine existing toolchains to generate customized workflows. The fulfillment of long-tail demands will spawn numerous research directions that were previously impossible.
Prediction 3: The Primary Workforce for Drug Design Will Shift from Human Experts to 24/7 AIDD Agents
Agents possess the ability to accumulate knowledge. We will see more and more companies transforming their years of accumulated drug discovery best practices into agent skills, continuously iterating and upgrading them over time. The era of relying on experienced experts for drug design is becoming history, while 24/7 always-on AI agents are coming online.
But experts won't disappear—they will evolve. Future experts will no longer be executors but rather trainers and calibrators of agents. They will be responsible for breaking down intuitions like "why this molecule makes me uneasy" into rules that agents can understand, and for hitting the pause button when agents produce results that "look reasonable but feel wrong."
Prediction 4: Human Experts Will Work Alongside AIDD Agents, Using Taste and Judgment to Guide Drug Discovery Projects
As Terence Tao recently discussed in a blog interview, AI research excels in breadth—it generates many good ideas—but lacks depth, specifically the intuition to dig deep in a particular direction. Human experts can exactly compensate for this weakness. Future biotech companies may operate in a "centaur" model, where human experts provide taste and judgment while AIDD agents handle execution.
In drug discovery, "taste" means:
- The intuition to choose the molecule that "feels more druggable" among ten equally reasonable candidates
- The courage to judge "whether this target is worth continuing to invest in" based on mechanistic understanding when data is insufficient
These are not objective functions that algorithms can optimize, but value judgment frameworks internalized by humans through countless trials and errors. For at least the next 10 years, "human in the loop" will remain the norm—not because AI isn't powerful enough, but because the essence of drug discovery is making value judgments under uncertainty, which is precisely what humans excel at.
Prediction 5: AIDD Agents Will Enable Autonomous DMTA Cycles, Dramatically Compressing Cycle Times and Improving R&D Efficiency
Because agents can execute tasks over extended periods and learn from environmental feedback, they can continuously update their understanding of drug discovery from both failures and successes. Previously, we hoped that drug design challenges could be solved by continuously upgrading and training single-point models (like AlphaFold). Perhaps these challenges can now be addressed within the broader framework of agents.
Real-world drug discovery is an iterative DMTA (Design, Make, Test, Analysis) cycle. In our view, different stages of DMTA require different AIDD base models. Connecting the inputs and outputs of these different models, along with correct data interpretation, strategy formulation, and task execution, are capabilities that agents already possess—capabilities that single-point models like AlphaFold inherently lack.
Therefore, we can envision scenarios where agents autonomously conduct DMTA cycle iterations. Currently, due to the time required for synthesis (Make) and biological evaluation (Testing), a single DMTA cycle typically takes 4-6 months. Initially, agent involvement may dramatically shorten the Design and Analysis phases (from weeks to days or even hours), but cannot significantly compress the entire cycle.
With advances in automated synthesis (such as Emerald Cloud Lab), robotic experimental platforms, and organoid technologies, we have reason to believe that individual DMTA cycles could be compressed to less than one month. Research projects that previously could only iterate 2-3 times per year may soon iterate 12 times annually. New drug development efficiency will achieve a qualitative leap, and corresponding costs will decrease.
Prediction 6: The Organizational Structure of the Drug Discovery Industry Will Change, Moving from Regional Dispersion to Concentration
When AIDD agents significantly compress DMTA cycles, logistics time becomes glaringly conspicuous. Previously, compounds could be synthesized in Beijing, sent to Suzhou for in vitro testing, and then arranged for animal experiments in Shandong—cost was the primary consideration, and you went wherever the quote was cheapest.
In the AIDD agent era, saving logistics time may be more important than getting lower quotes. When design + analysis is compressed from weeks to hours, 2-3 days of inter-city logistics becomes unacceptable. Robotic experimental platforms and automated synthesis equipment require physical space and cannot be "cloud-deployed" like software. Agents need to adjust strategies immediately based on experimental results, requiring "design-synthesis-testing" to occur continuously in the same space-time.
We predict that highly concentrated clusters of biotech companies + CRO companies will emerge in certain regions. Design, chemical synthesis, and biological testing will occur within the same park. This is not a regression to the "pre-internet era," but rather a redefinition of the value of "local" through agent technology—when computation and decision-making can be globalized, physical execution must be extremely localized.
Conclusion
In summary, we are both excited and concerned about the arrival of this new AIDD agent era. We are excited because our generation has the fortunate opportunity to witness a productivity revolution in the drug discovery industry. We are concerned because in this new era, will our importance as humans diminish?
Perhaps the question itself provides the answer. Agents do not worry about their own value, do not contemplate the meaning of work late at night, and do not feel professional satisfaction when seeing patient feedback after medication. These "unnecessary" thoughts and feelings are precisely the underlying forces that drive drug discovery from "making medicine" to "saving lives."
We conclude with a famous quote from English writer Charles Dickens: "It was the best of times, it was the worst of times." The era of transformation triggered by AI brings not only destruction but also rebirth; it is both crisis and opportunity. What distinguishes destruction from rebirth, crisis from opportunity, is not the precision of algorithms, but what kind of future the people using the algorithms want to create.