The most valuable use of an agent swarm here is to discover the variables that make biological responses predictable, not merely generate more candidate geroprotectors.
That gives us a sharper research objective:
[
\boxed{
\text{Discover what changes the organism’s responses to interventions,}
\quad
\text{then learn how to change those response properties safely.}
}
]The important distinction is between moving a system into a better state and restoring its ability to reach, maintain, and recover better states. I would organize the entire program around distinguishing those two.
And there is a striking real example that illustrates both the promise and the danger of the elegant dynamical framing.
1. What the linked paper gives us, and what we still have to discover
Zhavoronkov and Mishra’s September 2026 paper supplies an intervention-oriented framework: latent biological states, state-dependent intervention effects, functional targets, safety constraints, and a restoration-cost objective. It explicitly acknowledges that it has not trained an engine that learns intervention vector fields from perturbational data. Its worked dynamical example and its intervention-ranking exercise use specified biological coordinates rather than experimentally discovering an underlying dynamical basis. Aging-US
That missing identification step is precisely where I would put your swarm.
Rather than treating the paper’s axes as the answer, give the swarm permission to replace them. “Mitochondrial dysfunction,” for example, might turn out to be too broad, while some combination of recovery latency, metabolic reserve, and dependence on a neighboring cell population might better predict interventions.
Also, the paper’s control-based age depends on the available interventions, costs, safety requirements, and time horizon. It is therefore an intervention-relative assessment, not an intrinsic coordinate of the organism. Aging-US
My proposed division of labor would be:
[
\underbrace{\text{The paper}}{\text{specifies the control problem}}
+
\underbrace{\text{axis-discovery swarm}}{\text{discovers what system we are controlling}}.
]Writing an optimization objective does not identify the system that must satisfy it.
2. A real case your swarm should study: two aging trajectories need not mean two competing attractors
The yeast-aging literature contains almost exactly the situation you warned about.
A 2020 study described divergent aging trajectories through interactions between chromatin-silencing and mitochondrial pathways. A subsequent 2023 study engineered a synthetic oscillator involving Sir2 and HAP regulation and reported an 82% increase in yeast replicative lifespan relative to its wild-type control. This is an unusually direct example of engineering temporal dynamics rather than simply increasing or decreasing a molecular component. PubMed Central (PMC)
But a March 2026 PNAS study challenged the interpretation of the natural system. Using single-cell measurements and targeted perturbations, its authors argued that rDNA instability and mitochondrial decline can proceed independently and concurrently. Their proposed explanation for distinct terminal phenotypes is competing hazards, rather than mutual inhibition: whichever process reaches catastrophic failure first determines the observed ending. pnas.org
Schematically, the alternatives are:
[
\begin{aligned}
\textbf{Interacting-fates model:}&
\quad \text{one process suppresses the other};\
\textbf{Competing-failures model:}&
\quad T_{\rm death}=\min(T_{\rm failure,1},T_{\rm failure,2}).
\end{aligned}
]These can produce superficially similar endpoint patterns while implying different interventions.
In the first, altering a coupling might redirect an aging trajectory. In the second, suppressing one failure process may simply reveal the other.
The newer study does not, by itself, negate the engineered oscillator’s reported effect. But my inference is that an effective engineered intervention does not uniquely establish the model of natural aging used to motivate it. PubMed Central (PMC)
This should become a founding benchmark for the swarm:
Can it distinguish genuine dynamical coupling from apparent coupling created by survival selection, measurement, and competing failures?
That is much more informative than asking it to draw a more elaborate aging landscape.
3. What should count as an “original dynamical axis”?
I would separate three things that are too easily conflated.
A state coordinate describes what condition the system is in: a reservoir, burden, activation state, or cell-population composition.
A dynamical property describes how it behaves: recovery time, amplification, memory, coupling strength, transition probability.
A control property describes what interventions can accomplish: susceptibility, accessible directions, safe transition costs, or dependence on treatment order.
The discovery target could belong to any of these categories. They do not have to form one orthogonal basis.
There may not be a unique original coordinate system
Consider a local model:
[
\dot x=Ax+Bu,\qquad y=Cx.
]Changing latent coordinates using (x’=Sx) gives:
[
A’=SAS^{-1},\qquad B’=SB,\qquad C’=CS^{-1}.
]But the observable response is unchanged:
[
C’e^{A’t}B’=Ce^{At}B.
]So two agents can discover apparently different latent axes while recovering the same input–output behavior.
Consequently, I would not reward an agent for insisting that its latent variable is the fundamental aging axis. I would reward it for discovering a reproducible response structure and identifying measurements or interventions that give that structure biological meaning.
A useful operational definition is:
Two biological histories belong to the same effective state only when they predict approximately the same future functional responses under the admissible interventions we care about.
That is a much better quotient than “these cells have similar expression profiles.”
The central hidden-variable test
Suppose two cells appear identical under your representation:
[
z_A \approx z_B.
]Apply the same intervention and obtain reproducibly different responses:
[
P(Y_{\rm future}\mid z_A,u)
\neq
P(Y_{\rm future}\mid z_B,u).
]Something relevant is missing: history, spatial context, damage, exposure, cell identity, or another variable.
This is not merely a philosophical concern. Work on single-cell dynamical inference establishes that snapshot distributions alone generally do not determine the underlying dynamics without additional assumptions. PubMed
The swarm’s most productive hunting ground is therefore unexplained response heterogeneity, not unexplained variance alone.
4. The axes I would ask the swarm to search for
These are candidate descriptors and hypotheses, not a claim that aging has ten universal dimensions.
Candidate axis What might actually be changing? Discriminating experiment or measurement Possible intervention implication Restoring-force strength Perturbations decay more slowly. Compare recovery trajectories after matched disturbances, across several amplitudes. Restore recovery mechanisms rather than merely lower the resting biomarker. Transient amplification A disturbance spreads or briefly grows even though eventual recovery remains possible. Measure the peak, timing, and direction of cross-channel responses. Reduce harmful cross-coupling or improve containment. Intervention susceptibility The same input produces a smaller, larger, or differently directed response. Match delivered exposure and target engagement, then compare downstream trajectories. Restore responsiveness before applying the main restorative intervention. Memory and hysteresis Present behavior depends on previous stress, mechanics, or treatment. Match present measurements but vary prior histories; apply a second challenge. Erase harmful memory or establish a beneficial response history. Resource reserve and competition Several repair processes compete for a limiting resource. Apply isolated and combined demands; measure response saturation and recovery. Sequence repair demands rather than activate everything simultaneously. Phase and timing coordination Components retain function but act at poorly coordinated times. Vary relative timing while holding average exposure approximately fixed. Restore temporal organization rather than increase component abundance. Spatial access and niche dependence Tissue architecture changes exposure, communication, or permissible responses. Compare isolated cells, controlled matrices, and multicellular contexts. Repair the environment or delivery route, not only the cell. Task-relevant response repertoire Fewer distinct useful responses remain accessible. Compare the diversity of functional trajectories elicited by diverse perturbations. Restore missing capabilities rather than maximize nonspecific activity. Noise orientation and temporal structure Fluctuations increasingly point toward dangerous states, or persist longer. Separate perturbation responses from residual fluctuations and their correlations. Change buffering or noise propagation, not simply reduce all variance. Independent failure reserves Several damage processes advance independently toward distinct limits. Track them within the same units and perturb each selectively before terminal selection. Address whichever remaining failure process limits the benefit of another treatment. The important thing is that these hypotheses prescribe different measurements and different treatments.
“Old cells recover poorly” is not sufficiently specific. A cell could recover poorly because it has weak negative feedback, excessive transient amplification, inaccessible repair machinery, insufficient resources, or a persistent stress memory.
A useful swarm must keep those explanations separate until experiments distinguish them.
Why your pseudospectra interest belongs here
Here is a simple mathematical counterexample:
[
A_0=
\begin{pmatrix}
-1&0\
0&-2
\end{pmatrix},
\qquad
A_k=
\begin{pmatrix}
-1&k\
0&-2
\end{pmatrix}.
]Both have exactly the same eigenvalues. But starting from (x(0)=(0,1)), the second produces:
[
x_1(t)=k(e^{-t}-e^{-2t}),
]whose peak is (k/4).
Thus, the eventual decay rates can remain unchanged while transient cross-system effects become much larger. This is the general non-normal-dynamics distinction established in work on hydrodynamic stability; whether it explains a particular aging phenotype requires biological evidence. Science
An eigenvalue-only aging model could miss the relevant deterioration entirely.
Conversely, amplification is not automatically pathological. The question is whether it amplifies a useful response or pushes function toward a safety boundary.
5. Build the swarm around incompatible models, not merely different biological specialties
A swarm with an “epigenetics agent,” a “mitochondria agent,” and an “inflammation agent” may simply reproduce the existing ontology.
I would instead give agents different commitments about what kind of dynamical object the organism is.
Agent family Its alternative explanation What it must contribute Minimal-state agent A few variables suffice; extra structure is unnecessary. The strongest simple state-space baseline. Kinetic-mode agent Slow transitions and metastable states organize behavior. Predictive modes, transition times, and state-dependent switching. Transient-response agent Coupling and amplification matter more than slow eigenvalues. Impulse responses, transient gains, and input–output sensitivities. Memory agent Apparently inconsistent responses arise from omitted history. Delay states or memory models that improve held-out predictions. Mechanochemical/spatial agent Missing variables reside in mechanics, transport, or the niche. A specific spatial measurement that should resolve a predictive failure. Population/selection agent Apparent rejuvenation or branching is caused by changing composition, proliferation, or survival. Lineage-aware and competing-risk explanations. Intervention-composition agent Inputs change each other’s subsequent effects. Quantitative predictions for sequence, spacing, and combination experiments. Measurement/identifiability adversary The claimed axis is a batch effect, coordinate artifact, exposure difference, or non-identifiable parameter. The simplest alternative that reproduces the observation. These are not just debating personas. Each should produce an executable model or a precisely specified statistical test.
There are existing methodological building blocks. Controlled sparse equation discovery can search for compact dynamics with inputs; VAMP searches for predictive dynamical features; delay-coordinate methods can expose history-dependent structure; balanced model reduction connects representation to observability and controllability. None of these methods, by itself, guarantees a correct biological model. arXiv
The swarm’s advantage would come from coordinating these different failure modes, not from making every agent fit the same latent neural model.
Give every proposed axis a discovery contract
Each submission should specify:
Proposed variable → evidence → rival explanation → quantitative response prediction → additional measurement → falsifying result → intervention consequence.
For example:
“A latent history-dependent mechanical state explains why cells with similar present transcriptional profiles respond differently to regenerative stimulation. Measuring mechanical history and nuclear signaling should improve out-of-sample response prediction. If exposure, present-state measurements, and composition already explain the difference, the proposed memory axis is unnecessary.”
This forces “interesting idea” to become “distinguishable hypothesis.”
6. The discovery loop: observations, disagreement, then experiments
A. Start from a response atlas, not an age atlas
For a perturbation (u_j), estimate something like:
[
R_j(\tau\mid s)\mathbb E[y(t+\tau)\mid s,\operatorname{do}(u_j)]
\mathbb E[y(t+\tau)\mid s,\operatorname{do}(0)],
]where (s) denotes the measured starting context.
The atlas should index starting state, intervention, elapsed time, history, biological replicate, and functional outcome.
Factorizing this response array could reveal candidate response modes. But those modes are only discovery aids. Their apparent dimensionality depends on which interventions and sensors were included, how quantities were scaled, and how well small responses were measured.
The crucial question is:
Which candidate modes predict a held-out intervention or explain a previously unexplained failure?
B. Search jointly over representation, dynamics, and measurement
Do not freeze the latent space and only optimize the controller.
Let the swarm propose changes to:
[
\boxed{
\text{observables}
+
\text{state representation}
+
\text{memory}
+
\text{dynamics}
+
\text{intervention model}.
}
]Suppose expression-based models repeatedly fail only in stiff matrices. One branch might add mechanical state. Another might explain the failure through altered exposure. A third might attribute it to cell-population changes.
The next task is not to average those stories. It is to choose the measurement that distinguishes them.
This is where an agent can be genuinely useful as a designer of measurements: discovering that an extra time-resolved mechanical or functional measurement is worth more than another large static molecular panel.
C. Preserve diversity in predicted experiments
Your qualified-diversity idea becomes very concrete here.
Two models that use different terminology but predict the same responses are largely redundant. Two models that fit the existing data equally well but predict opposite responses to a feasible intervention represent valuable uncertainty.
I would maintain diversity across:
[
\text{predicted outcomes of candidate experiments},
]not merely across text embeddings, pathway names, or model architectures.
Agents should develop models independently before exchanging results. Afterward, merge genuinely redundant explanations while retaining experimentally distinguishable ones.
D. Spend experiments where models disagree
Choose the next experiment by its expected ability to distinguish the surviving models, subject to cost and safety constraints.
Optimal experimental design for discriminating biological models is an established research area; the swarm would extend that approach to a changing collection of representations and mechanisms. PubMed Central (PMC)
Importantly, preserve some exploratory measurements outside the current models’ preferred designs. Otherwise, the swarm could become exceptionally efficient at distinguishing several versions of the same mistaken premise.
An experiment that invalidates every model is a successful discovery step.
E. Require simple baselines to lose fairly
Compare against age-only predictors, static biomarkers, linear dynamical models, additive perturbation models, and single-agent analyses under matched budgets.
This is essential rather than ceremonial: a 2025 benchmark found that the tested deep-learning perturbation predictors did not outperform deliberately simple baselines on its single- and double-perturbation tasks. That does not settle every future benchmark, but it establishes why sophistication cannot serve as evidence of predictive advantage. Nature
7. Three particularly promising hypothesis programs
Hypothesis program A: aging changes responsiveness before it destroys capacity
Proposed mechanism: a repair capability still exists, but its activation becomes inaccessible or poorly coordinated.
In a local description,
[
\dot x=A(x,m)x+B(x,m)u,
]the relevant change could be in (B), the state-dependent response to an intervention, rather than solely in the uncontrolled dynamics.
A preliminary intervention might therefore be valuable because it restores the effectiveness of a second intervention.
Mechanical memory provides an empirical reason to take hidden history seriously: experiments in human mesenchymal stem cells found that prior exposure to stiff environments affected subsequent signaling and fate responses on softer substrates. That supports the existence of history-dependent mechanical state, not the broader claim that it is a universal aging axis. Nature
Swarm-generated prediction: cells matched on present molecular measurements will show different responses depending on mechanical or stress history; adding the relevant history variable will predict which preliminary intervention restores responsiveness.
Critical alternatives: different delivered exposure, different cell composition, or incomplete measurement of present state.
What would make this a treatment hypothesis: an intervention improves later functional responsiveness after its immediate effect has washed out, rather than merely changing a marker during treatment.
Hypothesis program B: aging increases noisy variation while reducing useful response diversity
This is a particularly important possibility for your diversity framework.
Proposed mechanism: cells become more variable at rest but less capable of producing distinct, appropriate responses when challenged.
That would give:
[
\text{resting heterogeneity}\uparrow,
\qquad
\text{task-relevant response repertoire}\downarrow.
]The swarm would compare the independent functional responses elicited by a varied intervention panel. It would ask whether some responses disappear, become indistinguishable, or become coupled to unacceptable side effects.
Discriminating prediction: a beneficial intervention could restore several distinct functional responses without increasing overall resting variance or substantially changing a static age score.
Falsifiers: the apparent loss disappears after correcting for sensor noise, delivered exposure, cell composition, or unequal sampling.
The therapeutic objective is not “maximize dimensionality.” A system that responds indiscriminately to everything is not healthy. The target would be selective access to useful capabilities.
Hypothesis program C: the important variable is time spent in damaging states, not average pathway activity
The yeast oscillator work makes this worth investigating, although generalizing from yeast to mammalian aging remains a hypothesis. The engineered system maintained changing dynamics rather than simply fixing both regulators at constant levels. PubMed Central (PMC)
A possible damage model is:
[
D(T)=\int_0^T q(x(t)),dt,
]where (q) becomes large after prolonged occupancy of particular states. A richer version would make damage depend explicitly on uninterrupted residence time or prior history.
Then two schedules with similar average molecular activity could produce different damage accumulation.
Swarm-generated prediction: temporal spacing, duty cycle, or phase relationships improve function beyond what average exposure predicts.
Critical controls: matched cumulative exposure, target engagement, recovery intervals, cell-cycle effects, and selection.
This reframes intervention discovery from:
“Which level is youthful?”
to:
“Which temporal patterns avoid commitment to damaging states while preserving necessary activity?”
8. Where Lie brackets, flow matching, and topology should enter
They should enter after identification has earned them a role.
Intervention order: useful, but not automatically evidence of geometric control
The linked paper emphasizes noncommuting intervention effects and Lie brackets. Aging-US
In an idealized smooth model, a difference between short (A)-then-(B) and (B)-then-(A) applications can reveal how the corresponding fields fail to commute.
But biological order effects could also arise from exposure persistence, toxicity, delayed target engagement, proliferation, or omitted memory.
So I would ask the swarm to predict the magnitude and timing of sequence effects, not merely announce that order matters. It must include the background dynamics and pharmacological state where relevant.
Also, a direction appearing in a mathematical Lie algebra is not automatically safely reachable. Real interventions have bounds, delays, and usually no freely available reverse operation.
Flow matching and Schrödinger bridges: propose paths, do not certify mechanisms
These can be useful candidate trajectory generators. But selecting a smooth or low-cost path between observed distributions does not establish that the biology follows that path or that an available intervention can implement it.
The general snapshot-identifiability problem remains. PubMed
My proposed role would be:
[
\text{generate candidate trajectory}
\rightarrow
\text{test against identified dynamics}
\rightarrow
\text{design discriminating intervention}.
]Topology: characterize experimentally supported possibilities
I would use topology on a directed, intervention-labeled transition structure.
A node would represent a reproducible functional regime. An edge would represent a transition supported under a specified intervention, context, time horizon, uncertainty, and safety requirement.
Then ask:
Does an intervention create a previously unavailable safe route?
Do apparently distinct routes collapse once hidden state is measured?
Does a bottleneck persist across measurement choices and model classes?Sheaf-like local-to-global methods could help organize incompatible tissue-specific models, but an inconsistency must first be tested against calibration errors and mismatched observations. It is not automatically a biological obstruction.
This is topology used to formulate testable questions, not topology used to decorate a latent embedding.
9. What I would actually build first for ScienceClaw
I would make the first project “Aging Axis Discovery: Distinguishing State, Dynamics, and Selection.”
First: a synthetic identification benchmark
Construct systems whose apparent decline looks similar but has different causes: slowing recovery, transient amplification, shrinking intervention susceptibility, hidden memory, and independent competing failures.
Hide some state variables and give every method the same measurement and experimental budget.
The swarm must recover which experiments distinguish the mechanisms, not merely fit the trajectories. This gives known-ground-truth tests before anyone interprets a learned human “aging operator.”
Second: the yeast model-discrimination benchmark
Use the contrasting aging models above as a biological test case. The 2026 study provides public single-cell RNA-sequencing data, including accession GSE318875, and reports longitudinal imaging and genetic perturbations. Reusable trajectory-level data and metadata should be audited before designing the benchmark. PNAS
Require agents to predict held-out perturbation responses and explicitly model survival selection. Do not reward them for repeating whichever paper’s interpretation they encountered first.
Third: one primary-cell or tissue system with repeated functional challenges
For a mammalian pilot, I would choose a system where the same biological units can be followed over time, interventions can be controlled, and a functional outcome can be measured.
The design should compare present-state models against models incorporating history, environment, and response dynamics. It should include withdrawal and rechallenge, not just treatment followed by one endpoint.
Large perturbation resources can help generate candidate mechanisms. For example, scPerturb harmonized 44 public perturbation-response datasets, and genome-scale Perturb-seq provides broad perturbational coverage. But those resources should not be mistaken for complete longitudinal maps of aging dynamics. Nature
These would be laboratory research programs, not self-directed treatment schedules.
The acceptance criterion
I would promote a proposed axis only when it:
improves held-out response prediction, survives plausible measurement and selection explanations, and changes a useful experimental or intervention decision.
A strong result could be:
“Adding one history-dependent state variable explains an otherwise unpredictable treatment response and correctly predicts which intervention sequence restores function.”
Another could be:
“What looked like an aging bifurcation is better explained by independent failures plus selective survival.”
The second is just as valuable. It prevents a beautiful but wrong controller from being built.
The central synthesis
The swarm should search for the distinctions biology forces us to retain.
That combines both sides of our elegance discussion:
[
\underbrace{\text{compress}}{\text{discard genuinely irrelevant variation}}
\quad\longleftrightarrow\quad
\underbrace{\text{enrich}}{\text{restore hidden state when predictions fail}}.
]Too much compression gives a beautifully simple aging clock that cannot predict interventions. Too much enrichment gives an enormous model that explains everything retrospectively and nothing prospectively.
The right discovery loop is:
[
\boxed{
\text{Find a response the model cannot explain}
\rightarrow
\text{propose the missing state or mechanism}
\rightarrow
\text{design the experiment that distinguishes it}
\rightarrow
\text{retain only what earns predictive necessity}.
}
]That is how an agent swarm could discover original dynamical axes: not by deciding in advance what aging ought to look like, but by finding what must be represented for its interventions to become predictably effective.