Prologue

A note on timing

Agents can now carry out multi-step professional work over several days, testing and correcting their own results along the way. As this capability enters organizations, we must rethink how people divide work, who makes decisions, and how learning happens. CataX is exploring this new form of organization, starting with biotech.

We acquire and develop real therapeutic assets and take them into clinical validation. Scientific judgment and agents’ ability to execute come together in one team, shortening the path from a therapeutic hypothesis to a clinical answer.

Drug development puts an organization’s capabilities to a real test: knowledge must become studies we can execute, and those studies must produce evidence strong enough to change a decision. This manifesto starts with biotech to explore how organizations in the agent age form intent, work with agents, take responsibility, and use experience to improve themselves.

The next medicine begins with what humans and agents learn from this one.

The Shock

The frontier moved into work we thought was uniquely human

Clinical development depends on rigorous reasoning: assembling evidence, testing assumptions, and turning questions into studies. Advances in agent-based mathematics and software engineering are changing the division of professional work between people and machines.

In 2025, Gemini Deep Think solved 5 of the 6 problems at the International Mathematical Olympiad, reaching IMO gold-medal standard.[1] In the months that followed, agents such as Aletheia entered research-level mathematics, proposing, checking, and revising proofs.[2]

In September 2026, Anthropic announced that Claude, working largely autonomously on the Prove2Me platform with occasional high-level guidance from researchers, had completed the Lean formalization of Fermat’s Last Theorem in 11 days. It proved 30,300 intermediate theorems, of which 29,500 were used in the final proof.[3] That October, OpenAI released more than 700 mathematical manuscripts generated by internal models; their validity remains under assessment by the mathematics community.[4] Production is beginning to outpace understanding.

Anthropic also put 16 agents to work together, building a C compiler capable of compiling the Linux kernel in 2 weeks.[5] People set the goal and built the testing and feedback environment. Agents implemented, debugged, and refined the code, bringing a complex system into being.

Agents have moved from generating content to sustained reasoning and hypothesis testing under strict constraints and verification. Clinical development also has to connect mechanism, study design, and evidence into a testable chain of reasoning, with answers supplied by experiments and clinical studies. As agents take on more of this reasoning, people must redefine their own work: choosing the questions worth pursuing, designing ways to test them, and taking responsibility for the resulting actions.

The first boundaries to move were the ones we thought were fixed.

The Biotech Curse

The cost of turning a promising molecule into a medicine

10–15 years of development. An average capitalized R&D cost of about $2.6 billion per approved new drug.[6][7] A promising molecule must cross this threshold of time and capital before it can become a treatment available to patients.

In 2012, Scannell and colleagues described the preceding decades’ predicament as Eroom’s Law: the number of new drugs approved per $1 billion of R&D spending had roughly halved every 9 years.[8] Better scientific tools had not automatically made the development system more productive.

AI drug discovery has already advanced molecular design. Antibodies designed with RFdiffusion have reached experimental screening and structural confirmation. Rentosertib, discovered and designed with the help of AI, was the subject of a randomized, placebo-controlled Phase 2a study published in 2025.[9][10] As candidates emerge faster, the development system must become better at testing them.

The clinical stage accounts for about 70% of direct preapproval R&D spending.[7] Whether a candidate benefits people depends on its mechanism, exposure, patient selection, and study endpoints. The next gain in productivity must reach these development decisions: turning richer knowledge into studies that can distinguish competing explanations, and using limited clinical resources to resolve the most consequential unknowns.

This is where CataX begins. We identify preclinical and early clinical assets worth developing, secure rights through license-in or co-development agreements, lead development through clinical proof of concept (PoC), and work with partners to take the medicines further. The mechanistic rationale, product differentiation, and questions an early study can answer determine where we invest our effort and capital.

A shelved asset brings the development question into sharp focus: does this molecule still have clinical value worth testing? A mechanism may have failed in the original population or at the original dose. It may never have received the resources needed for a proper test. Redevelopment begins by finding, in the existing evidence, the critical unknown that determines the next investment—and turning it into a question a study can answer.

Development efficiency ultimately rests on the quality of the answers. When a meaningful translational or patient-subgroup signal emerges, we design the next test. When results do not support further investment, we stop early and establish which hypotheses have been ruled out and which remain untested. Time and capital go toward evidence that can change a development decision.

A better development system changes which medicines become possible.

Rebuilding the Organization

New capabilities call for new forms of organization

In the early days of factory electrification, electric motors replaced steam engines, but machines and processes remained arranged around a central drive shaft. Only when factories removed the shaft and reorganized the floor did the new source of power truly change productivity.[11]

Today, assigning an agent to every medical, clinical, and operations role runs into the same problem. Individual steps get faster while the project still waits for handoffs between departments. The division of work once reflected the limits of people’s time and knowledge. As those limits shift, organizations must reconsider which tasks to separate and which questions one team must understand from beginning to end.

Base44 was founded by one person and had just 8 employees when it was acquired.[12][13] Teams like this show what a smaller group can now take on: a person who understands the problem can organize agents to handle more of the work from requirements to delivery.

Software can receive feedback soon after deployment. Clinical PoC in biotech often takes years. A small clinical development team must preserve the full scientific rationale, carry it through study design and execution, and remain accountable after the results arrive. A change in team size matters only if it changes how this work gets done.

An early decision made on a faulty premise can shape years of subsequent investment. An organization cannot wait for an efficacy readout to start learning. Before a study ends, prior experience, day-to-day feedback, and deliberate reasoning must already be changing the next decision.

Start with the whole question. Establish who is accountable for the result and which people and agents will contribute. Then define roles and divide the work.

What would we build if agents were there from day one?

Agentic from Day Zero

Agents as team members, infrastructure as the first hire

An agent-native organization builds the conditions for intelligence to take part in work from the outset. What knowledge can be accessed, which tools can be used, what authority is granted, and how results are tested form its basic operating structure.

Context. The reasoning behind a decision must survive changes in personnel. Source material, decision rationales, failure records, version histories, and unresolved questions together form organizational memory. The people and agents who follow can reconstruct the original reasoning and use it to develop new explanations.

Tools. Literature, databases, analytical code, and clinical data systems enter the workflow through APIs or controlled tools. Before patient data are connected, the purpose of use, scope of authorization, and privacy protections must be defined. Access is limited to the data required for the task, with de-identification, controlled processing, and records of access and export in place.[14][15]

Authority. Agents act within defined permissions and refer decisions beyond those permissions to the person responsible. Clinical trial sponsors and investigators retain their respective responsibilities.[14] Every piece of work that informs a regulatory submission or a patient-safety judgment is signed off by a named person who takes responsibility. Agents expand what that person can see and handle; they do not replace the signature. Quality reviews, release decisions, and critical changes are carried out and documented by authorized personnel within the quality system.

Evaluation. Agent tasks have test cases, quality standards, and failure criteria. Before use, systems supporting GxP work must undergo computerized system validation (CSV) appropriate to their intended use and risk. Electronic records and signatures subject to 21 CFR Part 11 must have the required access controls and audit trails; system changes are assessed, validated, and approved.[15][16] Performance on real tasks determines whether models and workflows have improved.

Build the infrastructure that lets every lesson become a capability.

Self-Overcoming

Scientific intent sets the direction

When complex execution can be entrusted to agents, professional identity needs a new foundation. Checking informed consent forms, reformatting protocols, and reconciling CRO spreadsheets consume large amounts of professional time. As that burden eases, the harder work is deciding what is worth doing and what evidence would justify a change of direction.

As execution gets faster, taste, scientific intuition, and the quality of the questions become scarcer. Scientific intuition helps us spot anomalies worth pursuing in messy evidence. Taste shows in the choices we make: which study deserves the resources, and which polished analysis has missed the point.

Intent turns human ambition into a direction for work: what to change, why it matters, and what result would count as success. In drug development, the asset lead uses that intent to formulate a scientific claim, organize agents to seek evidence and counterexamples, and turn the claim into a testable study. The more we can do, the more clearly we must understand what we are pursuing.

Nietzsche described humanity as a bridge.[17] Self-overcoming here is concrete work: stepping beyond a role we have mastered, posing our own questions, bringing resources together, and taking responsibility for the outcome. Growth also means letting go of answers that once served us and rethinking a problem when new evidence arrives.

We must grow with the power we gain.

Beyond Job Titles

From functional handoffs to ownership of the whole question

A drug’s development hypothesis spans mechanism, exposure, patients, and clinical endpoints. BD acquires an asset on the strength of its mechanism. The clinical team turns the thesis into a protocol and executes it. If the original reasoning is lost along the way, a failed PoC leaves the team struggling to tell whether the drug did not work or the study never truly tested the hypothesis.

Build the asset team around the whole development question: in which patients, and at what exposure, can this drug deliver enough clinical value to justify further development? The same human–agent team takes responsibility from asset assessment through study design and evidence generation. Agents maintain the source material, reasoning, and change history. The lead ensures that every trade-off can be traced to the hypothesis being tested.

Team members collaborate through agreed data interfaces, delivery standards, and feedback mechanisms. Deviations during execution and new findings return to the asset team’s shared context. One team remains responsible for the question posed at the start, the changes made along the way, and what the results actually answer. Tools and methods developed through each program enter organizational memory, giving the next asset a stronger starting point.

Titles follow capability. Structure follows the work.

Recursive Organization Improvement

The organization itself must become the thing that improves

Experience becomes an organizational capability only when it changes subsequent work. The tools, reasoning, and ways of collaborating developed during a program must be available for the people and agents who follow to use, test, and revise. In our earlier work at CataX, we called this joint improvement of people, agents, knowledge, tools, and workflows Recursive Organization Improvement.[18]

Learning Actively Before Results Arrive

Answers about clinical value take years. Organizational learning cannot stand still. Data-driven learning and rationale distillation are two paths that test and refine each other. The first finds problems in real work and project outcomes. The second is led by people who actively construct, distill, and challenge the logic of development.

For a candidate with a compelling mechanism, the discussion cannot stop at “the target is validated.” Why these patients? Can the expected exposure produce the required effect on the target? Can the study endpoint test this chain of reasoning? What finding would change the investment decision? The lead organizes literature, historical cases, and professional experience around these questions. Agents compare the evidence and seek counterexamples, turning a promising mechanism into a development hypothesis we can act on.

Routine checks, rework, and execution deviations provide fast feedback to improve tools and processes. Mechanistic reasoning is first tested against historical cases and questions from other disciplines, then corrected by experimental and clinical results. This approach adds checkpoints between sparse outcome signals, turning everyday execution and tests of reasoning into feedback that guides the next action.

Codifying Judgment into Architecture

Effective methods are encoded in Skills; the sequence of collaboration and the relationships between deliverables are encoded in Workflows. Organizational memory retains why these methods work, where they apply, and how they have failed. The next team of people and agents can use established approaches and see when they no longer fit.

The people who understand the problem best also help improve the tools. In biotech, an asset lead turns repeated document checks into a program; a clinical lead builds checks for common omissions into a tool. After testing and expert review, the method enters a shared library of capabilities. Users then refine it against real problems.

Evaluating the Evaluator

Directed evolution carries useful changes into the next round through variation, selection, and accumulation.[19] Working methods can undergo the same kind of selection. People and agents propose alternatives; task performance and project outcomes determine which are worth keeping.

Evaluation criteria must themselves be tested against outcomes. If an evaluation rewards a complete report but misses a risk that determines an asset’s value, we change the criteria and rerun the comparison against the cases we have accumulated. When clinical results expose blind spots in earlier reasoning, they must also change the questions we ask when choosing the next asset and designing the next study.

Many companies can access the same models. Project experience has to be earned firsthand, with the infrastructure to capture and build on it. Once that experience is embedded in tools and workflows, each stronger generation of models can use it to take on more complex work.

Every program should improve the organization that develops the next.

A Recipe

Nine ingredients for a self-improving organization

Building an agent-native organization means bringing expanded capabilities, human responsibility, and accumulated experience together. These 9 ingredients form a recipe we can keep refining in practice.

1. Own the Decision

   Keep decisions and accountability for their consequences together. A named person explains why resources are being committed and what would change the decision, then takes responsibility for what follows.

2. Build for Agents from Day Zero

   From day one, build organizational memory, tools, and work records that agents can use, so new members can pick up where others left off.

3. Give Every Human a Team

   Put a team of agents in the hands of someone who understands the problem, extending expertise into action. Give people room to set goals, draw on resources, and revise their methods.

4. Turn Repetition into Capability

   When a task comes up for the 3rd time, extract what can be reused. Put tested methods into shared tools, Skills, and Workflows, and retain the conditions under which they apply.

5. Evaluate Before Deployment

   Evaluate agent tasks for their intended use before deployment. Test, review, and authorize significant changes, then track performance in practice.

6. Organize Around the Question

   Form teams around the whole question. Use delivery agreements to coordinate specialist work, and maintain accountability from the initial question through to the result.

7. Make Failure Useful

   Record the rationale behind a failed decision, the causes of error, and the specific changes made. Carry those lessons into tools and processes, and check whether similar problems become less frequent.

8. Improve How We Improve

   Combine project data with human-led rationale distillation to test how we find problems, evaluate outcomes, and retain improvements.

9. Measure Meaningful Progress

   Put real outcomes ahead of process metrics. In biotech, measure the quality of evidence, the time it takes to obtain it, and the capital consumed. Calibrate those measures against subsequent clinical results.

Good ideas earn their place in practice.

The Call

To builders, inside and outside biotech

CataX starts with biotech, putting this approach to organization to work in real drug development. Every trade-off ultimately bears on whether a treatment can move forward and when patients gain a new option.

We welcome partners, scientists, and builders who share our long-term conviction. You have depth in a field and are willing to understand the questions next to it, form your own views, and organize people and agents to do the work. You can build tools for real problems and revise your methods when new evidence arrives.

We work with asset owners to advance real medicines and test our ways of organizing through actual programs. Work in materials, energy, semiconductors, and law also faces long validation cycles, consequential decisions, or regulatory constraints. We welcome builders in these fields to exchange practices and turn the experience of one undertaking into capability for the next.

Self-Improving Biotech points to a broader question: when intelligence can be organized and experience passed on, what kinds of organizations can we build—and what kinds of builders can we become? CataX is seeking the answer in real programs.

Patients are waiting for treatments. We answer with results.