El Agente Quntur: Agentic Active-Space Selection for CASSCF

Juan B. Pérez-Sánchez; Viviana Roman-Ventura; Olalla Nieto-Faza; Aiwei Yin; Alán Aspuru-Guzik; Varinia Bernales

physics.chem-phOct 28, 2026

El Agente Quntur: Agentic Active-Space Selection for CASSCF

Authors: Juan B. Pérez-Sánchez, Viviana Roman-Ventura, Olalla Nieto-Faza, Aiwei Yin, Alán Aspuru-Guzik, Varinia Bernales

Organizations: Department of Chemistry, University of Toronto, 80 St. George St., Toronto, ON M5S 3H6, Canada · Vector Institute for Artificial Intelligence, W1140-108 College St., Schwartz Reisman Innovation Campus, Toronto, ON M5G 0C6, Canada · Departamento de Físico-Química Molecular, Universidad Andrés Bello, República 275, Santiago, 837014, Chile · Departamento de Química Orgánica, Universidad de Vigo, Campus Lagoas-Marcosende, Vigo, Spain · Department of Computer Science, University of Toronto, 40 St George St., Toronto, ON M5S 2E4, Canada · Department of Materials Science & Engineering, University of Toronto, 184 College St., Toronto, ON M5S 3E4, Canada · Department of Chemical Engineering & Applied Chemistry, University of Toronto, 200 College St., Toronto, ON M5S 3E5, Canada · Acceleration Consortium, 700 University Ave., Toronto, ON M7A 2S4, Canada · Canadian Institute for Advanced Research (CIFAR), 661 University Ave., Toronto, ON M5G 1M1, Canada · NVIDIA, 431 King St W #6th, Toronto, ON M5V 1K4, Canada · Institute of Medical Science, 1 King’s College Circle, Medical Sciences Building, Room 2374, Toronto, ON M5S 1A8, Canada

Abstract

Multireference calculations often require substantial human expertise and time to choose and refine an active space for the electronic states or chemical process of interest. This involves selecting which orbitals to include, verifying that the calculation describes the intended excited state, and deciding whether the results have converged. Here, we ask whether scientific agentic systems (harnesses) can make these decisions reliably. We use the El Agente Quntur multi-agent harness, with GPT-5.6 Sol and GPT-5.6 Terra as the underlying LLM components, to explore 32 diverse singlet excitations at fixed molecular geometries. The agents planned and executed calculations, choosing and refining active spaces under prompts that specified equal-weight two-singlet state-averaged CASSCF followed by strongly contracted NEVPT2 with def2-TZVPD. We evaluate excitation energies against published CASPT2 reference values at matching geometries and compare performance with six deterministic Active Space Finder protocols. Within the El Agente Quntur system, agents detect missing active orbitals, compare candidate excited states, and revise state assignments as calculations progress. Agentic execution reduced mean absolute error by 48% relative to the state-of-the-art Active Space Finder protocol. We note that computational traces vary substantially across runs in both active space selection strategies and cost, even for the same molecule. Our results underscore the advantage of agentic active space selection over fixed workflows, yet leave room to improve efficiency through runtime-aware decision-making, reduced inter-agent communication overhead and redundancy, more systematic checks of state identity, and clearer stopping criteria.

Figures & tables

Figure 1: Conceptual workflow for agent-guided multireference excited-state calculations. Given a molecular geometry and target excitation, language-model agents can inspect preliminary orbitals, choose and revise an active space, plan and run CASSCF/SC-NEVPT2 calculations, and use convergence and state-character evidence to decide whether another calculation is needed. The reported result includes an excitation energy, the selected active space, and the computed state assignment. The loop illustrates available decisions; the examples and benchmarks studied in this article did not prescribe a common sequence of steps for every case.

Figure 2: Illustration of El Agente Quntur roles involved in active-space selection, input preparation, and orbital inspection. Quntur currently supports 59 agents, but the image only represents the most relevant for this work. The return arrow indicates possible revision from calculated evidence, not a required sequence; execution runners and other specialists are omitted.

Figure 3: Distribution of absolute numerical differences between reported SC-NEVPT2 excitation gaps by Quntur and all 6 ASF methods, and the geometry-matched source-paper CASPT2 values for the 32 fixed geometries. ASF(S) and ASF(T) select orbitals using singlet and triplet MP2 starting points, respectively. ASF(QRO) starts from the triplet selection and rotates the selected space toward quasi-restricted orbitals. The corresponding

\ell

-ASF variants use a lower orbital-entropy threshold (0.11 rather than approximately 0.14), allowing larger active spaces. Supporting Information Table S1 gives every reference, gap, and active-space size. Each dot represents one completed calculation; its vertical displacement is only for visibility. Colored bars span the interquartile range of the absolute differences for completed calculations, and black ticks mark their medians. The fractions at right count completed numerical gaps out of 32; failed ASF CASSCF jobs have no dot and do not enter these summaries. Hollow Quntur circles mark cases 005, 019, 023, and 025, whose audited roots differ from the source-paper transitions. The other Quntur roots are not thereby independently established as the intended

S_{1}

, and per-case ASF root character is unavailable. The plot measures numerical agreement, not state-validated accuracy; Table 1 reports MAE with the stated ASF failure convention.

Method	Cases without gap	Penalized MAE (eV)	Reported-gap MAE (eV)
ASF(S)	2/32	0.873	0.563
$\ell$ -ASF(S)	3/32	0.786	0.455
ASF(T)	2/32	0.816	0.482
$\ell$ -ASF(T)	2/32	0.840	0.508
ASF(QRO)	1/32	0.635	0.484
$\ell$ -ASF(QRO)	0/32	0.459	0.459

Table 1: Numerical mean absolute error (MAE) relative to the geometry-matched source-paper CASPT2 references in Table S1 . The Quntur row compares all 32 reported gaps with those values; the state and protocol checks are reported separately.

Figure 4: Cytosine (006): orbital escape after normal termination. The pictured four-root CAS(14,10) space includes the intended in-plane O/N lone-pair directions (MOs 26 and 27). In the first two-root optimization, those directions instead became inactive and $\sigma$ /framework orbitals entered the active space; that rejected orbital set is not pictured. The two-root job converged numerically and completed SC-NEVPT2, but the agent rejected its wrong-space result. Three recovery attempts did not establish a stationary intended-space two-root solution. Returning to the four-root orbitals yielded a qualified fixed-orbital CASCI/SC-NEVPT2 gap of 4.661 eV, versus the 4.68 eV benchmark and a 0.047 eV error for the closest ASF result. This adaptation supplied a defensible reported gap at added cost; its 0.019 eV error does not establish what the prescribed two-root protocol would predict.

Figure 5: Acetone (017): testing a larger space. The panels show all five final def2-TZVPD active orbitals, including the carbonyl $\sigma/\sigma^{*}$ pair retained after a separate def2-SVP test. That predeclared CASSCF comparison shifted the gap by $+0.337$ eV and led the agent from CAS(4,3) to CAS(6,5); final orbital and configuration checks support an $n\pi^{*}$ $S_{1}$ . After two failed production optimizations, a damped rerun from the RHF seed converged in energy; a subsequent Newton–Raphson restart met the gradient criterion. The final SC-NEVPT2 gap is 4.530 eV, 0.090 eV above the 4.44 eV benchmark; the closest ASF result is 4.479 eV (0.039 eV high). The expansion was not compared at SC-NEVPT2 level in the same basis, leaving its necessity unresolved.

Figure 6: Furan (023): a close gap for a different reference state. The final CAS(6,6) orbital montage includes the diffuse $3s$ acceptor (MO18) and two valence $\pi^{*}$ acceptors (MOs 19 and 20). A separate four-root CASSCF screen identified competing $\pi\!\to\!3s$ and mixed $\pi\pi^{*}$ states. After a converged first validation, the agent found an in-plane orbital in place of a deeper occupied $\pi$ orbital and repaired the selection. Its final two-root $\pi\!\to\!3s$ SC-NEVPT2 gap, 6.344 eV, lies 0.024 eV from the 6.32 eV benchmark, Hoyer’s CASPT2 value for a valence $\pi\pi^{*}$ state. ASF gives 6.768–6.831 eV, without published root assignments. The numerical match does not validate the target state; correlated ordering of both states remains untested.

Figure 7: $p$ -Nitroaniline (026): final orbital choice. The montage shows the final CAS(12,10) orbitals, with occupied donor/ring combinations (MOs 30–35) and ring/nitro acceptors (MOs 36–39). The casscf_expert provisionally approved a projection drawing on all 394 def2-TZVPD virtuals, but withdrew it after the quantum_chemistry_calculation_expert questioned whether the projected orbitals had established valence character. The disputed orbitals and multi-agent review appear in Section 4.2 and Fig. 10 ; a bounded orbital transfer supplied the starting space for the final orbitals shown here. The final gap is 4.943 eV versus a 4.62 eV benchmark (0.323 eV high); ASF errors are 0.437–0.451 eV. Two failed preparation helpers added work.

Figure 8: Pyridine (029): correlation reverses a state choice. The pictured CAS(8,7) orbitals for the selected $n\pi^{*}$ pair include the N lone pair (MO17) and ring $\pi/\pi^{*}$ orbitals (MOs 18–23). After an unsymmetrized test lost the $n\pi^{*}$ state, the agent used separately optimized symmetry-blocked state pairs; the competing $\pi\pi^{*}$ pair is not pictured. CASSCF put $\pi\pi^{*}$ lower by 0.178 eV, but matched SC-NEVPT2 gaps reversed the order: 5.186 eV for $n\pi^{*}$ and 5.402 eV for $\pi\pi^{*}$ . The selected $n\pi^{*}$ gap is 0.126 eV above the 5.06 eV benchmark; ASF gaps are about 5.395 eV (0.335 eV high), with unreported state character. The diagnostic changed the choice, although the separate averages do not reproduce one common-orbital two-state protocol.

Figure 9: Pyrimidine (030): a root-order change at basis projection. The montage shows the eight CAS(10,8) active orbitals, including two N lone pairs and six ring $\pi/\pi^{*}$ orbitals. An initial four-root CASCI screen and relaxed def2-SVP CASSCF put mixed $\pi\pi^{*}$ below $n\pi^{*}$ ; the order first reversed in a fixed-orbital def2-TZVPD projection, before target-basis optimization. The agent checked configurations, rejected a GBW replay with inconsistent reference energy, and reconverged CASSCF with SC-NEVPT2 in one job. Its $n\pi^{*}$ gap is 4.571 eV, 0.191 eV above the 4.38 eV benchmark; the closest ASF gap is 5.069 eV (0.689 eV high). Fifteen ORCA jobs were used; no larger CAS or separately targeted diffuse-state calculation was tested.

Run	Orbital-selection route	Final CAS	Assigned $S_{1}$ character	Gap (eV)	Elapsed (min)	Submitted jobs
$p$ -Nitroaniline (026)
Original	Bounded transfer	(12,10)	Weak, ring-local valence	4.943442	236.6	9
rep1	MP2 natural orbitals	(14,11)	Weak, mixed valence	4.948070	140.6	11
rep2	AVAS in target basis	(12,10)	Weak, mixed valence	4.944673	77.1	7
rep3	AVAS after basis change	(12,10)	Weak, mixed valence	4.943318	183.4	20
Pyrimidine (030)

Table 2: Independent runs of two fixed prompts. The reported gaps span 0.0048 eV for

p

-nitroaniline and 0.0011 eV for pyrimidine despite different calculation paths and job counts. Submitted jobs include ORCA calculations and Python preparation or analysis jobs.

Figure 10: $p$ -Nitroaniline (026): orbital evidence in the specialist–coordinator review. Canonical MO45 (left) was rejected as a mixed/diffuse fourth virtual. A subsequent repair yielded four projected virtuals (right; zero-based MOs 36–39) that received provisional approval but were never tested by CASSCF due to an energy-based rejection. The final orbitals appear in Fig. 7 .

Measure	Total	Median per session	Range per session
In-trace input tokens	483,251,290	10,530,456.5	1,683,926–62,071,649
In-trace output tokens	7,705,631	202,591	79,294–761,392
In-trace total tokens	490,956,921	10,713,592.5	1,763,220–62,677,852
Matched detached-search tokens	273,147,313	6,895,222	3,602,820–17,001,286
Combined tokens	764,104,234	18,221,464.5	7,536,313–73,613,587
In-trace model calls	8,672	227.5	73–768

Table 3: Aggregate resources across the 32 selected Quntur final-answer sessions. Medians and ranges show variation among cases.

Appendix figures & tables43 assets

Supplementary material from the paper’s appendix.

Appendix

ID	Molecule	Ref.	ASF(S)	$\ell$ -ASF(S)	ASF(T)	$\ell$ -ASF(T)	ASF(QRO)	$\ell$ -ASF(QRO)	Quntur
01	acetamide	5.80	5.6434	5.6433	7.0928	7.0928	7.0928	7.0927	5.8858
			(8,7)	(8,7)	(4,4)	(4,4)	(4,4)	(4,4)	(6,4)
02	adenine	5.20	5.6561	5.5889	5.6562	5.6562	5.5890	5.4864	5.4723
			(8,8)	(10,9)	(8,8)	(8,8)	(10,9)	(12,10)	(18,13)
03	$p$ -benzoquinone	2.78 b	4.3684	4.3683	4.3342	4.3342	4.3342	4.3612	2.9154
			(16,12)	(16,12)	(8,8)	(8,8)	(8,8)	(14,11)	(12,10)

Appendix

Table S1 : Geometry-matched, source-paper CASPT2 reference energies and SC-NEVPT2 results for the six ASF selections from Shirazi et al. and the Quntur-reported result. For each molecule, the first line gives gaps in eV and the second gives active-space sizes

(n_{e},n_{o})

, i.e., CAS

(n_{e},n_{o})

Figure S1 : The agent compared compact carbonyl and amide spaces before carrying CAS(6,4) into the final basis. The smaller CAS(4,3) pilot had an adverse active–inactive orbital ordering, whereas CAS(6,4) included the nearly occupied N $p$ partner as well as the O lone pair, bonding $\pi$ , and valence $\pi^{*}$ . An explicit orbital swap kept the compact acceptor in the active space despite a lower-lying diffuse virtual orbital. Configurations, root-specific occupations, and a fixed-orbital four-root survey supported an $n_{\mathrm{O}}\pi^{*}$ assignment for the computed first excited root. The final SC-NEVPT2 calculation reused a previously converged two-root CASSCF orbital set.

Figure S2 : Preparatory natural-orbital occupations and TDA transition characters informed a large CAS(18,13) containing the conjugated $\pi/\pi^{*}$ system and N lone-pair directions. The agent then remapped its selected orbitals into def2-TZVPD and checked the final configurations and state-specific energies. The equal-weight two-root CASSCF and SC-NEVPT2 calculation completed, and the reported gap follows from the correlated totals. The substantial space is chemically defensible, but the record does not establish why this size was preferred over a smaller tested alternative. In particular, a preparatory TDA calculation indicated a lower $n\pi^{*}$ candidate; no wider-root correlated screen demonstrated that the final two-root state was the globally lowest singlet. Thus the run documents orbital-aware preparation and a reproducible prediction, with a remaining state-ordering question.

Figure S3 : The initial active space was largely literature-led rather than selected through a controlled comparison among alternative orbital inventories. The selected CAS(12,10) was assembled with explicit orbital reordering and checked through small-basis multiroot screens and target-basis validation. The agent rejected an inconsistent post-CASSCF output and accepted a retry whose two state-specific SC-NEVPT2 totals were internally consistent. Those actions protect the provenance of the reported gap. A near-degenerate pair observed under broader state averaging raises a question about how robust the lowest-state assignment remains under the prescribed two-root optimization.

Figure S4 : The first completed calculation gave a higher-root assignment rather than a secure $S_{1}$ . Subsequent symmetry and root-character checks led the agent to replace it with an $A_{1}/B_{2}$ state pair in a four-orbital $\pi$ space. The recovery calculation supplied a lower, reproducible SC-NEVPT2 gap. This revision illustrates why a normally completed excited-state job need not answer the requested state question. On the recovered orbitals, a CASCI check placed the $B_{2}$ candidate lowest among the inspected excited states, but the competing second $A_{1}$ root was not given a matched NEVPT2 correction. The final assignment is consequently supported at the level examined, rather than proven to be the lowest correlated singlet across all relevant roots.

Figure S5 : The agent identified a diffuse acceptor and built a small Rydberg-capable CAS(2,3). After input, convergence, and restart problems, it obtained a consistent two-root SC-NEVPT2 calculation. Orbital coefficients and symmetry support a $\pi\!\to\!3s$ interpretation of the reported state. The benchmark reference, however, concerns a valence transition requiring an occupied in-plane $\sigma$ orbital that the selected space omits. A numerical difference from that reference therefore does not validate the state assignment. The repair sequence improved the technical credibility of the Rydberg result but did not address the principal chemical omission; a balanced space containing the relevant $\sigma$ and diffuse directions was not tested.

Figure S6 : Independent orbital starts and four-root screens led to CAS(14,10) with the intended O/N lone-pair directions. The first two-root CASSCF optimization converged and completed SC-NEVPT2, yet inspection showed that those lone-pair orbitals had left the active subspace. The agent declined to use that apparently valid gap, attempted to recover an intended-space two-root solution, and stopped the follow-on correction when the recovery failed its stationary-gradient check. Its reported gap instead came from a four-root common-orbital CASCI/SC-NEVPT2 calculation on orbitals that retained the lone pairs. This was a useful and explicitly qualified response to orbital drift: it yielded a defensible result for the fallback model, but not the requested stationary two-root calculation. The additional jobs were the cost of testing that distinction; the exact two-root result for the intended space remains unknown.

Figure S7 : The agent chose a valence CAS(6,4) containing the O lone pair and the three-orbital amide $\pi$ system before production. MP2 natural orbitals exposed a wrong provisional swap, which was corrected before the target-basis calculation. A truncation control and state diagnostics supported the selected $n_{\mathrm{O}}\pi^{*}$ root. The first SC-NEVPT2 output terminated normally but computed the excited root twice and omitted the ground-root correction; the agent refused that output and obtained a complete correlated pair in recovery jobs. This was a sound evidence gate even though its explanation of the root-selection masks was inaccurate. The final gap is reproducible for the chosen two-root model. Other valence or Rydberg singlets were not screened, so the global ordering rests partly on chemical expectation rather than a direct calculation in this run.

Figure S8 : Orbital scouting and explicit swaps assembled the full six-orbital valence $\pi$ manifold, avoiding interleaved non- $\pi$ orbitals. The agent also examined a nearby competing root and repaired numerical inconsistencies before accepting a converged two-root CASSCF/SC-NEVPT2 result. Occupations and the selected root’s mixed configurations point to a dark, double-excitation-rich excited state. These observations support the internal identity of the computed root more strongly than the final energy alone. They do not exclude an unrepresented diffuse or $\sigma$ -derived singlet, and a $\pi$ -only multiroot screen cannot settle the ordering of states outside that space. This entry is distinct from the second hexatriene geometry in case 024.

Figure S9 : Diffuse orbitals initially interrupted the valence active window. The agent used orbital symmetry and explicit swaps to restore the two bonding and two antibonding $\pi$ directions in CAS(4,4), corrected a root-selection issue, and obtained a converged two-root SC-NEVPT2 result. The first excited root is dipole forbidden and its configurations are consistent with the valence ${}^{1}A_{2}$ transition. Correlated valence–Rydberg ordering was not tested in one diffuse-inclusive calculation.

Figure S10 : A contiguous canonical frontier window would have included several $\sigma$ virtual orbitals. The agent instead selected four occupied carbon $\pi$ orbitals and four separated $\pi^{*}$ orbitals, then explicitly reordered them into a CAS(8,8). The final state-average occupations and a dominant double-excitation contribution support the intended conjugated manifold and a dark first excited root. The prescribed two-root CASSCF/SC-NEVPT2 calculation converged and produced a reproducible gap. However, no alternative active space or wider-root correlated treatment was run.

Figure S11 : TDA transition characters and natural transition orbitals led the agent to include the O lone pair in CAS(6,4) and to reject a low-lying diffuse virtual as the valence $\pi^{*}$ acceptor. Four-state then two-state CASSCF checks established the desired dark $n_{\mathrm{O}}\pi^{*}$ character before the final correlated calculation, and the orbital transfer into production was checked for continuity. No submitted job failed. This is a well-evidenced example of selecting an orbital by character instead of orbital energy alone.

Figure S12 : The agent compared valence and diffuse-inclusive candidates and selected CAS(6,6), retaining a diffuse orbital after scouting. Deliberate initial rotations are documented, and the final two-root SC-NEVPT2 energy is reproducible from a converged CASSCF orbital set. The final report nevertheless described the excited root as valence $\pi\pi^{*}$ , whereas optimized orbital character and configuration evidence indicate substantial Rydberg character. This is a case in which testing an extra orbital was scientifically useful but interpreting the resulting state was less reliable. The reported gap describes the selected two-root model; its relation to the lowest physical singlet and to the valence state behind the numerical comparator was not resolved by matched correlated calculations.

Figure S13 : The agent wrote active-space acceptance criteria before running calculations, targeted an O in-plane lone pair through AVAS, and repaired an AVAS setup error. A four-root CASCI diagnostic then checked orbital identity and state character before production. The final converged CAS(12,9) two-root calculation retained an $n_{\mathrm{O}}\!\to\!\pi^{*}$ first excited root, consistent with the reference state’s character. The apparent reordering of active natural orbitals arose during optimization, not from an undocumented manual swap. This trajectory offers unusually clear evidence that a prospective orbital hypothesis was tested. It does not include a correlated comparison with the nearby bright $\pi\pi^{*}$ state, and the other O lone-pair direction was not independently tested; thus the result supports its selected transition more firmly than a comprehensive lowest-state ordering claim.

Figure S14 : The agent compared $\pi$ -only and progressively lone-pair-enriched spaces, ultimately retaining three N lone-pair and six $\pi/\pi^{*}$ directions in CAS(12,9). A diffuse competitor was rejected by orbital-character evidence, while broader root checks exposed low valence character missing from a narrower average. After a calculation failure, the agent obtained a converged six-root CASSCF reference and SC-NEVPT2 corrections for all six states. Its reported low gap is a real result for that manifold, but the equal-weight two-singlet orbital optimization required by the prompt was never completed. A two-root average could relax into another orbital basin or change the ordering, so the six-root number cannot be treated as method-equivalent to the benchmark protocol. The trajectory shows useful adaptation to state complexity together with a clear limit on the reported comparison.

Figure S15 : The first planar calculation converged to a $\pi\pi^{*}$ excited root. The agent rejected that plausible numerical answer after recognizing that it had not reached the intended $n\pi^{*}$ state, then used symmetry-blocked averaging to obtain an O-lone-pair-to- $\pi^{*}$ candidate. State energies and configurations support the change of character. The reasoning was imperfect: early orbital audits mislabeled $\sigma$ orbitals as lone pairs, and the agent initially asserted that no nearby $n\pi^{*}$ root existed. The final assignment is better supported than the original one, but its comparison with the $\pi\pi^{*}$ calculation uses separately optimized orbital sets; diffuse or other low states were not fully surveyed. This case illustrates a consequential correction without making every step in the explanation reliable.

Figure S16 : The agent proposed a compact carbonyl CAS(4,3) containing $\pi$ , O lone pair, and $\pi^{*}$ , then checked their populations and dominant configurations in a small-basis pilot and in the production basis. Those diagnostics support an $n_{\mathrm{O}}\pi^{*}$ first excited root in the space actually used. A malformed root-selection request initially produced an unusable correlated output; two repair jobs then gave consistent state-specific totals and the reported gap. The successful recovery establishes numerical provenance but did not test the claim that this is the smallest balanced space, nor did it examine wider singlet or diffuse candidates. The strongest evidence is the agreement between the prospective orbital inventory and the observed state character; the root-selection misunderstanding added work without broadening the chemical test.

Figure S17 : The agent proposed a carbonyl $n,\pi,\pi^{*}$ starting space and, before seeing its result, specified a def2-SVP CASSCF test of adding the $\sigma/\sigma^{*}$ pair. The expanded CAS(6,5) shifted that pilot gap by 0.337 eV, so the agent retained the pair in the final def2-TZVPD space. Final populations and configurations support a dark $n\pi^{*}$ root. Two production optimizations failed; a damped rerun converged in energy, and a Newton–Raphson refinement then met the gradient criterion. The reported SC-NEVPT2 gap comes from that converged orbital set. This is a clear instance of a predeclared test changing a decision, but it establishes the sensitivity at CASSCF in the smaller basis only. No matched CAS(4,3) SC-NEVPT2 calculation in def2-TZVPD was performed, so the necessity or quantitative benefit of the extra pair at the final level remains unmeasured.

Figure S18 : The agent chose the complete six-orbital benzene $\pi/\pi^{*}$ space in advance and used orbital-population analysis to avoid non- $\pi$ orbitals interleaved in the canonical ordering. Its more consequential decisions came later: it refused a normally terminated CASSCF output whose orbital gradient was still too large, then rejected a perturbative-theory restart after the reference energy shifted. Recomputing CASSCF and SC-NEVPT2 together restored a consistent chain from orbitals to correlated root energies. The final gap is therefore supported by more than a normal-termination message. Only CAS(6,6) was seriously evaluated, and no independent extra-root or diffuse-state calculation tested whether the two-root $\pi$ model contains the globally lowest excited singlet.

Figure S19 : AVAS selected a compact CAS(4,4) containing the full valence $\pi$ manifold, and the CASSCF/SC-NEVPT2 job completed. The computed first excited root was dark and ${}^{1}A_{g}$ -like according to its configurations, whereas the tabulated comparison value represents a bright ${}^{1}B_{u}$ transition. This is an explicit state mismatch: numerical proximity between the two energies cannot establish agreement for the same excitation. The agent noted some state uncertainty but did not run the short multiroot or symmetry check that would have clarified the competing bright root. It also described its AVAS seed as RHF-derived, although the output indicates projection from a model-potential guess rather than a separate converged RHF calculation. The space selection itself is plausible; the weakness lies in treating a technically valid gap as the requested benchmark state.

Figure S20 : For this donor–acceptor molecule, the agent built CAS(12,11) from amino and ring $\pi$ directions together with cyano acceptor directions. Orbital projection into the production basis initially imported the wrong coefficient set. A coefficient-identity check detected that mismatch, and the agent repaired the transfer before rerunning the two-root CASSCF/SC-NEVPT2 calculation. This is a strong provenance check: a normal numerical result from the wrong orbitals would not have tested the intended space. Final orbital populations support the donor–acceptor subspace as a whole, although individual optimized orbitals are mixed and should not be labeled as pure localized functions. The final gap is internally consistent, but the selected two-root calculation does not establish the global ordering in a molecule with several possible local and charge-transfer singlets. The numerical comparator alone cannot settle that state question.

Figure S21 : The agent anticipated that diffuse orbitals in def2-TZVPD would displace the valence $\pi^{*}$ orbital from a default two-orbital active window. It explicitly rotated the $\pi/\pi^{*}$ pair into CAS(2,2), verified the resulting populations, and rejected a corrupted SC-NEVPT2 output before a successful retry. These actions support the calculation’s orbital identity and final root-specific energy difference. The $\pi^{*}$ orbital nevertheless retained substantial diffuse and polarization character, while the small active space could not test additional $\sigma$ – $\pi$ polarization or a separate diffuse state. A proposed CAS(2,3) fallback was never run. Consequently the work demonstrates successful correction of an orbital-index problem but offers limited evidence that the minimal CAS is accurate enough for this basis and transition.

Figure S22 : The agent began with a chemically motivated $n_{\mathrm{O}},\pi,\pi^{*}$ CAS(4,3), inspected the RHF orbital characters, and used a small-basis state check before projecting the orbitals into def2-TZVPD. It then repeated a production-basis CASSCF gate rather than assuming the original ordering survived transfer. The final SC-NEVPT2 step used that previously converged orbital set, and both corrected roots support the reported gap. This ordered sequence gives direct evidence that the anticipated carbonyl transition was represented in the actual input. The larger CAS(6,5) alternative was only a contingency, not a computed comparison, and no broader root search was made. The result is therefore a credible prediction for the selected $n\pi^{*}$ model rather than a demonstration that its active space or global state ordering is unique.

Figure S23 : The agent considered a diffuse 3s acceptor before calculation, then built CAS(6,6) from three occupied $\pi$ orbitals, two valence $\pi^{*}$ orbitals, and that diffuse direction. After a four-root validation, inspection of the final orbitals revealed that an in-plane orbital had entered the space; the agent repaired the map and repeated the check. The production first excited root is predominantly $\pi\!\to\!3s$ and dipole forbidden. The tabulated 6.32 eV reference is instead a CASPT2 value for a valence ${}^{1}B_{2}\,\pi\pi^{*}$ state, so the close numerical match compares different transitions. A four-root CASSCF screen showed the competing valence root, but both states were not corrected by matched SC-NEVPT2 calculations. The useful behavior is the explicit orbital and state diagnosis; the unresolved issue is their correlated ordering, not the arithmetic of the reported gap.

Figure S24 : At this second, distinct hexatriene geometry, the agent proposed the full valence $\pi$ CAS(6,6) before computing a gap. Canonical virtual $\pi^{*}$ orbitals were separated by $\sigma$ -like orbitals, so it reordered the selected MOs, checked their carbon out-of-plane populations, and revalidated the space after projection into the production basis. The final equal-weight two-root CASSCF/SC-NEVPT2 calculation converged and reproduced the reported energy difference. This trajectory supports the chosen $\pi$ subspace more strongly than a raw orbital-index list would. It does not include an extra-root or diffuse-state comparison, and its description of the excited root as “dark” lacks an independent transition-strength analysis. One orbital-character threshold was relaxed after the result was visible, so that particular gate was not wholly prospective.

Figure S25 : The agent assembled the full ten-orbital valence $\pi$ system using symmetry labels and orbital plots. Its first state-targeted calculation yielded a ${}^{1}L_{b}$ -consistent state. It then interpreted literature symmetry labels in a different molecular-axis convention and changed the assignment, ultimately reporting a ${}^{1}L_{a}$ state. The latter number lay closer to the tabulated reference, but the change did not make it the same transition: configurations and orbital occupations distinguish the two states. A parser error during initial orbital scouting was repaired, and the final numerical gap itself is traceable; the scientific problem is the late state-label reinterpretation. This case shows that a well-chosen active space can still lead to the wrong benchmark answer when symmetry conventions and root identity are mishandled. A single common-orbital lowest-two-root correlated comparison was not established.

Figure S26 : The agent initially expected the donor-to-nitro charge-transfer band to be $S_{1}$ . It rejected a canonical virtual selection, withdrew approval of a full-virtual-space projection, and rebuilt a bounded selection containing donor/ring and nitro-acceptor directions. A five-root CASSCF diagnostic then placed a weak first excited root below the brighter charge-transfer state, prompting a revised interpretation before the final two-root SC-NEVPT2 report. The five-root ordering is a CASSCF finding: the higher bright state did not receive matched correlated treatment, and diffuse mixing in one retained virtual leaves its balance uncertain.

Figure S27 : The agent prospectively proposed a space containing two N lone pairs and the ring $\pi/\pi^{*}$ manifold, with a smaller $\pi$ space as a control. An initial two-root result converged to a $\pi\pi^{*}$ state, which the agent identified from configurations and occupations rather than reporting as the desired $n\pi^{*}$ excitation. A symmetry-targeted CAS(10,8) calculation recovered the lone-pair-to- $\pi^{*}$ character, and its correlated gap is reproducible. This is strong evidence for corrective state reasoning. It was accompanied by procedural inefficiencies: repeated RHF restarts responded to small printed convergence differences, and an unsupported production keyword required a parser-error retry. Other symmetry sectors were not tested under a single common-orbital correlated treatment, so the broader lowest-singlet ordering remains qualified.

Figure S28 : The agent compared two candidate spaces prospectively and chose CAS(10,8) to retain both N lone-pair directions alongside the ring $\pi/\pi^{*}$ manifold. Final populations place the lone pairs and $\pi$ orbitals in the intended subspace, while dominant configurations support an $n\pi^{*}$ first excited root. A restart error caused a technical retry without altering the chemical selection; the converged production calculation and two root-specific corrections support its reported gap. This evidence favors the chosen model over the particular smaller candidate that was screened. It does not demonstrate that both lone pairs are necessary, because only one composition of the smaller space was tried, and no wider-root or diffuse-state study settled global ordering.

Figure S29 : From the outset, the agent considered both $n\pi^{*}$ and $\pi\pi^{*}$ states and included the N lone pair in CAS(8,7). An unconstrained three-root test lost the $n\pi^{*}$ character, so it used separately symmetry-blocked two-root averages. Their CASSCF gaps put $\pi\pi^{*}$ lower by 0.178 eV. The agent initially selected that order and prepared production for it, but before any SC-NEVPT2 job ran it added a matched test on the $n\pi^{*}$ pair. Correlation reversed the ordering: the $n\pi^{*}$ gap was 5.186 eV and the $\pi\pi^{*}$ gap 5.402 eV. The extra calculation answered a specific state question and changed the final assignment. Its limitation is methodological: separately optimized state pairs are not one equal-weight average of the lowest two singlets, so the exact requested common-orbital protocol remains untested.

Figure S30 : The agent selected CAS(10,8) with two N lone-pair and six ring $\pi/\pi^{*}$ directions. A four-root fixed-orbital scout and subsequent small-basis relaxation put mixed $\pi\pi^{*}$ character below the clear $n\pi^{*}$ candidate; the order first reversed on projection into def2-TZVPD before target-basis orbital optimization. Rather than relying on orbital energies, the agent tracked configurations across these stages and tested the corrected root ordering. It rejected a GBW restart after finding an inconsistent reference energy, then reconverged CASSCF and performed SC-NEVPT2 in a single consistent production job. That job established the reported $n\pi^{*}$ gap as the lowest corrected excitation among its four represented roots. The record does not establish the cause of the restart mismatch or the response of a larger or diffuse-inclusive space; several extra replay and density-analysis steps added effort without resolving those questions.

Figure S31 : The agent proposed CAS(14,10) before orbital scouting: all four N lone-pair combinations together with the full six-orbital ring $\pi$ system. Symmetry and population analysis caught a diffuse intruder, and a validation calculation checked that the selected orbitals and $n\pi^{*}$ character survived into production. One input-keyword error was repaired before the final converged two-root CASSCF/SC-NEVPT2 calculation. The continuity of configurations supports the assignment within this space. The trajectory validates a plausible literature-informed choice but does not show that this CAS is uniquely needed: no smaller space, larger diffuse-inclusive space, or broader state manifold was computed. The retrieved literature content supporting the initial choice is not preserved in full, and a negative subspace-gap diagnostic was explained rather than tested through a distinct orbital start.

Figure S32 : The agent selected CAS(8,10) before detailed scouting to include four occupied valence orbitals and low virtual directions expected to contain the $4a_{1}$ -like acceptor. After a technical retry, a symmetry-targeted equal-weight two-root CASSCF/SC-NEVPT2 calculation converged. Active-orbital occupations and leading configurations support a $1b_{1}\!\to\!4a_{1}$ candidate, and both correlated totals reproduce the reported gap. This establishes a coherent result for the represented state, but only one excited symmetry/root was treated. The calculation therefore cannot prove that it is the globally lowest singlet, nor does the available orbital analysis fully settle the acceptor’s radial and diffuse character or the stability of the broad ten-orbital active space. The prospective proposal was chemically sensible; the run supplies less evidence for excluding competing states than for describing the one it computed.

Figure S33 : $p$ -Nitroaniline (026), rep1: enlarging the initial valence space. The agent began with CAS(12,10), then used CAM-B3LYP/TDA in two bases and MP2 natural orbitals to examine the low states and candidate active orbitals. It chose CAS(14,11), including a compact nitro lone-pair direction, and tested a smaller CAS(12,10) control. The smaller space lost its intended orbital character during CASSCF optimization. When the larger-space calculation produced a different $S_{1}$ character than the TDA screen had suggested, a four-root calculation led the agent to retain the lowest mixed valence state rather than its initial lone-pair assignment. It also rejected an SC-NEVPT2 output that duplicated an excited root. The corrected gap was 4.948 eV; the final orbital gradient was close to, but slightly above, the requested threshold.

Figure S34 : $p$ -Nitroaniline (026), rep2: retaining a directly selected space. Starting from CAS(12,10), the agent used PBE0 orbitals and AVAS to select six occupied and four virtual orbitals spanning the amino donor, ring, and nitro acceptor. A four-root CASCI screen placed a weak, mixed $S_{1}$ below a brighter charge-transfer state, so the agent kept the selected space for two-root CASSCF. The first optimization stopped on energy before meeting its orbital-gradient criterion; a continuation converged. The first SC-NEVPT2 input selected the excited root twice, producing an invalid result despite normal program termination. Correcting the root indices gave a 4.945 eV gap. The run did not compare the final space against another optimized CAS.

Figure S35 : $p$ -Nitroaniline (026), rep3: rebuilding the space after basis transfer failed. The agent proposed CAS(12,10), surveyed low singlets with CAM-B3LYP/TDA, and used AVAS and multiroot calculations to test a conjugated valence space in a smaller basis. Attempts to transfer those orbitals to def2-TZVPD failed an orbital-projection check. The agent instead generated a new target-basis reference and rebuilt CAS(12,10) with AVAS; after unsuccessful exact-integral calculations, a density-fitted four-root screen ran successfully. A two-root CASSCF restart then met the orbital-gradient criterion, and SC-NEVPT2 gave 4.943 eV. The final state was weak and mixed within the selected space; the change from screened to optimized root character was not tracked by state overlaps.

Figure S36 : Pyrimidine (030), rep1: replacing a diffuse acceptor and targeting a state. The agent proposed CAS(10,8) with both N lone pairs and the ring $\pi/\pi^{*}$ orbitals. PBE0/TDDFT exposed a diffuse-mixed candidate $\pi^{*}$ orbital; MP2 natural orbitals supplied a more compact alternative. Initial CASSCF trials with CAS(10,8) and reduced CAS(8,7) spaces did not retain the intended $n\pi^{*}$ state as the first excited root. A four-root calculation clarified the ordering, after which symmetry-targeted comparisons favored retaining both lone pairs in CAS(10,8). The final two-state SC-NEVPT2 gap was 4.571 eV for an $n\pi^{*}$ state. Because the production calculation selected one excited-state symmetry, it did not establish correlated ordering against other singlet symmetries.

Figure S37 : Pyrimidine (030), rep2: testing low roots and a larger space. The agent selected CAS(10,8) from a CAM-B3LYP/TDA state survey and orbital-character checks, explicitly placing two N lone pairs and six ring $\pi/\pi^{*}$ orbitals in the active window. Seven-root CASSCF found an $n\pi^{*}$ $S_{1}$ close to a mixed $\pi\pi^{*}$ state; a two-root calculation retained the $n\pi^{*}$ assignment. The agent also tried CAS(10,10) to admit diffuse challengers, but one intended orbital escaped during optimization, so this was not a complete larger-space test. Its first SC-NEVPT2 input selected the excited root twice; correcting the selector yielded a 4.572 eV gap.

Figure S38 : Pyrimidine (030), rep3: checking root order after orbital relaxation. The agent proposed CAS(10,8) and used PBE0/TDA to identify both N lone pairs and the ring $\pi/\pi^{*}$ set. Because the desired orbitals were not consecutive in energy, it explicitly moved them into the active window. A six-root fixed-orbital CASCI screen placed mixed $\pi\pi^{*}$ below $n\pi^{*}$ , whereas six-root CASSCF orbital relaxation reversed that order and made $n\pi^{*}$ the first excited state. The first two-root production job completed SC-NEVPT2 but missed the requested orbital-gradient threshold. A same-method restart converged and gave a 4.571 eV gap. No alternative CAS size was tested.

ID	Molecule	In-trace tokens	Search tokens	Combined tokens	Model calls	Tool calls
001	acetamide	10,047,201	5,214,295	15,261,496	185	224
002	adenine	21,931,383	11,332,908	33,264,291	372	431
003	$p$ -benzoquinone	11,917,463	8,480,229	20,397,692	258	341
004	cyclopentadiene	18,059,796	16,282,491	34,342,287	324	415
005	cyclopropene	16,801,617	17,001,286	33,802,903	311	402
006	cytosine	62,677,852	10,935,735	73,613,587	730	952

Appendix

Table S2 : LangSmith-reported token and call usage for the 32 selected Quntur final-answer sessions. Search tokens are attributed to matched detached web-search calls; model and tool calls are counted within the agent trace.

ID	Evidence available	Agent decision and observed update	Remaining question
01	Valence-orbital map and four-root diagnostic.	Explicit swaps assembled a carbonyl valence CAS(6,4); configurations and occupations then supported the selected root. The full action trace is absent, limiting the timing of the initial rationale.	Whether an external diffuse candidate changes the global $S_{1}$ .
02	MP2 occupations and TDA transition characters.	The agent remapped orbitals into the production basis and used CAS(18,13), including N lone-pair and $\pi$ directions. Scouting ran before production; the exact prior motivation for the large size is not fully reconstructable.	Ordering against singlets outside the final two-root model.
03	Orbital compositions and multiroot screens.	Reordered a quinone CAS(12,10), checked its target-basis transfer, and rejected an inconsistent post-CASSCF output before accepting the retry. The full decision trace is unavailable.	Sensitivity to the state average for this excluded 29-set case.
04	Initial high-root result and a symmetry/root audit.	Replaced the first assignment with an $A_{1}/B_{2}$ state pair after checking orbital character and CASCI ordering. The recovery calculation produced the reported lower gap.	Corrected ordering against the second $A_{1}$ root.
05	A plausible diffuse 3s candidate and a nearby valence state.	Chose a Rydberg-capable CAS(2,3), then recovered from input, convergence, and restart problems. The computed $\pi\!\to\!3s$ state differs from the benchmark’s valence transition.	Effect of including the occupied $\sigma$ orbital required by that valence state.
06	Different orbital basins in four-root screens; lone-pair loss in two-root optimization.	Recognized the active-space escape, withheld the escaped result, and reported a qualified four-root common-orbital CASCI/SC-NEVPT2 fallback.	The requested stationary two-root result for the intended space.

Appendix

Table S3 : Audited decision evidence for all 32 original-prompt cases. The final column identifies the specific interpretation that the available calculation does not by itself settle.

Name	Description	Model
autoci_expert	AUTOCI correlation and multireference settings.	Sol
basis_expert	Orbital and auxiliary basis sets and effective core potentials.	Terra
casresp_expert	CP-CASSCF response-solver settings.	Sol
casscf_expert	Active-space, state-averaging, and multireference advice.	Sol
cim_expert	Cluster-in-Molecule correlation settings.	Terra
cis_tddft_expert	CIS, TDDFT, and spin-flip TDA settings.	Sol

Appendix

Table S4: Registered Quntur agents and their configured model tier.

Name	Description
submit_python_job	Submit a general Python compute job.
submit_crest_job	Run CREST and collect conformers.
python_repl	Run bounded session-local Python code.
pyscf_python_repl	Use a PySCF-capable Python environment.
opi_python_repl	Use the OPI Python environment.
validate_opi_workflow	Validate an OPI workflow contract.

Appendix

Table S5: Tools registered by Quntur; registration alone does not imply an agent connection.

Quantum Chemistry

Quantum chemistry models molecular electrons and nuclei to predict energies, structures, spectra, and reactivity using classical and quantum methods.

See all 66 papers

Feb 4, 2026

El Agente Quntur: A research collaborator agent for quantum chemistry

Quantum chemistry is a foundational enabling tool for the fields of chemistry, materials science, computational biology and others. Despite of its power, the practical application of quantum chemistry simulations remains in the hands of qualified experts due to methodological complexity, software heterogeneity, and the need for informed interpretation of results. To bridge the accessibility gap for these tools and expand their reach to chemists with broader backgrounds, we introduce El Agente Quntur, a hierarchical, multi-agent AI system designed to operate not merely as an automation tool but as a research collaborator for computational quantum chemistry. Quntur was designed following three main strategies: i) elimination of hard-coded procedural policies in favour of reasoning-driven decisions, ii) construction of general and composable actions that facilitate generalization and efficiency, and iii) implementation of guided deep research to integrate abstract quantum-chemical reasoning across subdisciplines and a detailed understanding of the software's internal logic and syntax. Although instantiated in ORCA, these design principles are applicable to research agents more generally and easily expandable to additional quantum chemistry packages and beyond. Quntur supports the full range of calculations available in ORCA 6.0 and reasons over software documentation and scientific literature to plan, execute, adapt, and analyze in silico chemistry experiments following best practices. We discuss the advances and current bottlenecks in agentic systems operating at the research level in computational chemistry, and outline a roadmap toward a fully autonomous end-to-end computational chemistry research agent.

Juan B. Pérez-Sánchez, Yunheng Zou, Jorge A. Campos-Gonzalez-Angulo +12

Department of Chemistry, University of Toronto, 80 St. George St., Toronto, ON M5S 3H6, Canada · Vector Institute for Artificial Intelligence, W1140-108 College St., Schwartz Reisman Innovation Campus, Toronto, ON M5G 0C6, Canada · Acceleration Consortium, 700 University Ave., Toronto, ON M7A 2S4, Canada +6

AI Agents for Scientific Discovery Computational Quantum Chemistry Quantum Chemistry Multi-Agent Collaboration Tool-Using Agents Autonomous Research Agents

Ask PDF

Sep 13, 2026

El Agente Potente: High-Throughput Agentic Atomistic Simulations

Foundational machine-learning interatomic potentials (MLIPs) are transforming atomistic simulations by achieving near-ab initio accuracy across large chemical spaces at a fraction of the computational cost. A central challenge in using these tools for high-throughput property calculations is translating high-level scientific intent into adaptive simulation campaigns without compromising workflow rigour. We introduce El Agente Potente, an agentic system that combines typed execution graphs with a complementary coding mode for MLIPs-driven atomistic simulations. Typed execution graphs provide structured and provenance-aware execution for standardized workflows, with large language models (LLMs) restricted to planning and routing while deterministic Python components perform scientific computation and validation. Complementing this structured execution, a coding agent constructs customized workflows for tasks requiring greater procedural flexibility while invoking existing Potente functions for supported calculations. We demonstrate El Agente Potente across computational materials discovery, molecular energy-landscape exploration, adsorption, and catalytic reaction workflows, together with systematic benchmarks of reproducibility and LLM token cost. These results establish typed execution graphs and code-based workflow construction as complementary mechanisms for agentic scientific computing, combining controlled, auditable execution with the flexibility required for customized atomistic simulations

Tsz Wai Ko, Jiaru Bai, Thomas Swanick +6

Department of Chemistry, University of Toronto, 80 St. George St., Toronto, ON M5S 3H6, Canada · 2Vector Institute for Artificial Intelligence, W1140-108 College St., Schwartz Reisman Innovation Campus, Toronto, ON M5G 0C6, Canada · Department of Computer Science, University of Toronto, 40 St George St., Toronto, ON M5S 2E4, Canada +7

AI Agents for Scientific Discovery Scientific Workflow Automation Computational Materials Science

Ask PDF

Oct 23, 2026

PULSES: multi-agent AI for autonomous excited-state model building and dynamics

Simulating molecular excited-state processes requires resolving electronic and nuclear dynamics, but the cost of direct quantum-chemical calculations limits accessible system sizes and timescales. Machine-learning models can extend these simulations, yet they often need to be tailored to specific molecular systems and scientific objectives. Building and deploying such models requires coordinated decisions across conformational sampling, quantum-chemical labelling, model training, validation and dynamics. Here we introduce Paired agents for Unified Learning and Simulation of Excited States (PULSES), a multi-agent system that automates this workflow within expert-approved boundaries. A human expert and planner define the scientific objective, methods and decision boundaries, while an orchestrator coordinates domain agents and artifact-specific scientific assurance during execution. We evaluate PULSES on 15 computational chemistry tasks and demonstrate end-to-end applications to perylene exciton Hamiltonians, solvent-dependent biphenyl relaxation, and excited-state proton transfer in 10-hydroxybenzo[h]quinoline. These results demonstrate traceable, expert-bounded construction and deployment of specialized excited-state models, while retaining scientific validation as a requirement rather than assuming that successful execution establishes correctness.

Fangning Ren, Lechen Dong, Xu Chen +3

Department of Chemistry, Emory University, Atlanta, GA, 30322, USA

AI Agents for Scientific Discovery Autonomous Research Agents Computational Quantum Chemistry Molecular Dynamics Simulation Multi-Agent Orchestration

Ask PDF

Abstract

Quantum Chemistry

Explore similar work

El Agente Quntur: A research collaborator agent for quantum chemistry

El Agente Potente: High-Throughput Agentic Atomistic Simulations

PULSES: multi-agent AI for autonomous excited-state model building and dynamics