El Agente Gr\'afico: A Semantic Execution Runtime for Scientific Agents

1 Introduction

• Typed execution graphs govern admissible scientific state transitions while preserving object identity, native state, failures and provenance.The runtime turns agent-generated procedures into executable protocols with explicit types, decision boundaries and recovery paths.Source: §1 · p. 2
• The runtime keeps large intermediate objects in Python memory and persists selected scientific state and provenance in an external knowledge graph for reuse across tools and sessions.Each workflow is exposed as a graph tool, while the knowledge graph provides longer-term state reuse.Source: §1 · p. 2
• El Agente Gráfico is evaluated through benchmark comparison, chemistry workflows testing typed transfer and persistence, open-ended investigations, and skill-guided graph authoring.The evaluations span quantum chemistry, ensemble spectroscopy, MOF design and MOF literature mining.Source: §1 · p. 2
• A focused controller selects only graph-permitted successor nodes, supplies typed inputs, and receives validation errors for correction.Large scientific objects bypass the language-model context while node outputs update later routing decisions.Source: Figure 1 · p. 3

2.1 Typed execution graphs govern semantic decisions over scientific state

• Typed GPU4PySCF graphs constrained routing to admissible transitions while validating inputs and reusing large scientific objects as native Python state.The aqueous pKa example selected frequency calculation after optimization because Gibbs free energies required minimum verification and thermochemical corrections.Source: §2.1 · pp. 3–4, Figure 2 · p. 4

2.2 Externalized execution improves task performance and efficiency

• 90.94% rubric-based performance accompanied 80.3% lower model cost and 4.5-fold shorter wall-clock time than El Agente Q on the direct sonnet-3.7 comparison.Requests fell from 168.4 to 9.1 per task, while trace tokens fell from 1,649,616 to 284,036.Source: §2.2 · p. 5
• 0.99 pass@3 and 0.54 pass3 for gpt-5 separated the probability of at least one successful run from the probability that all three runs succeeded.A successful run required numerical score 1.00 and rubric score above 0.90.Source: §2.2 · p. 6

2.3.1 Typed state transfer enables parallel ensemble spectroscopy

• A 17 nm merocyanine absorption shift between water and n-heptane arose primarily from the solvent model and was only weakly altered by conformational averaging.Ensemble maxima were 412.8 nm in water and 429.9 nm in n-heptane.Source: §2.3.1 · pp. 6–7
• Explicit water shifted 2,3-epoxybutanol’s lowest bright absorption from 9.47 to 9.21 eV and its ensemble maximum from 121.8 to 130.9 nm.The workflow combined concurrent gas-phase sampling, water-shell construction, typed transfer and ensemble analysis.Source: §2.3.1 · p. 6

2.3.2 Persistent relational state supports iterative MOF design

• Persistent MOF entities enabled concurrent construction, cross-session querying and topology-aware analysis of pore size and accessible surface area.Changing topology with components fixed could decouple pore size from accessible surface area, while within-topology trends did not persist across topologies.Source: §2.3.2 · pp. 8–9

2.3.3 Bounded execution supports open-ended chemical investigations

• Bounded execution let the agent choose chemical scope and interpretation while graphs validated inputs, constrained calculations and recorded outputs.The disulfide study supported an E–C–E/E–C mechanism, whereas the olfaction study compared vibrational spectra without validating a vibrational mechanism of olfaction.Source: §2.3.3 · p. 9

2.4 From workflow execution to skill-guided protocol authoring

• A graph-construction skill enabled coding agents to separate LLM extraction from deterministic identity checks, structure curation and Zeo++ calculations.The workflow was developed on five papers and tested unchanged on three reserved papers.Source: §2.4 · p. 9, §2.4.1 · pp. 9–10

• Scientific validity still requires domain knowledge and empirical evidence beyond schema validation and runtime traces.Source: §3 · p. 11
• Externalizing execution makes structural and procedural validity explicit, reducing conversational coordination while separating active native objects from persisted state and provenance.Execution-graph topology further constrains which state transitions may follow one another.Source: §3 · p. 11
• Agent-authored graphs become a second verification target because different implementations, curation choices, and validation checks can change execution outcomes.Recommended revisions should use interface checks, domain invariants, regression tests, held-out cases, versioning, and rollback.Source: §3 · p. 11
• Distributed laboratory coordination will require operation scheduling, rate limiting, and scoped access to instrument state and scientific context.Source: §3 · p. 11

• Typed execution graphs manage heavyweight objects and deterministic validation, invoking the model only at semantic decision points.This division reduces context burden and leaves a traceable workflow record.Source: §4 · pp. 11–12
• 80% lower model cost and 4.5-fold shorter wall-clock time accompanied improved task performance across six university-level quantum chemistry exercises versus the previous multi-agent architecture.Source: §4 · p. 12
• Redox and olfaction studies used execution offloading while the agent defined each investigation’s chemical scope and computational pathway.Source: §4 · p. 12
• A reusable graph-construction skill enabled coding agents to author typed execution graphs for dynamic MOF literature mining and computational workflows.Source: §4 · p. 12

5.1 Gráfico infrastructure

• The infrastructure maps typed Python classes and relationships into a knowledge graph with strict validation, lazy loading, NumPy-array support, and agent-oriented graph authoring.Pydantic-based agents use typed tools, dependency injection, sandboxed execution, and retry handling.Source: §5.1 · p. 12, §5.2 · p. 13

5.2 Computational tools

• The computational stack evaluates quantum chemistry, molecular generation, conformers, MOF assembly, structure equivalence, topology, porosity, and geometry optimization through interoperable packages.ASE provides the common atomic-configuration representation, while RDFLib, SPARQLWrapper, and Blazegraph support graph queries and persistence.Source: §5.1 · p. 12, §5.2 · pp. 13–14

5.3 Graphical user interface

• The GUI streams agent responses and tool logs while rendering molecular structures, trajectories, plots, downloadable files, and dynamic model selection.Source: §5.3 · p. 15

A List of tools

• The MOF workflow dynamically chains download, decompose, search, build, optimize, and analyze stages according to user intent.It supports database refcodes, local CIF files, direct PORMAKE construction, and knowledge-graph searches for compatible combinations.Source: A List of tools · p. 25

Parameters schema

• The documented tool interfaces expose structured schemas for MOF workflows, quantum chemistry, knowledge-graph access, unit conversion, molecular visualization, conformer searches, cluster growth, and Python execution.These schemas specify required identifiers, optional calculation settings, allowed units or methods, and execution constraints such as read-only SPARQL queries and timeouts.Source: Parameters schema · pp. 26, 29, 30

B Typed execution graph abstraction

• The abstraction layer combines shared workflow state, typed routing configurations, and type-derived node classes while remaining independent of chemistry-specific logic.Domain workflows construct specialized graphs by subclassing these components on top of pydantic-graph.Source: B Typed execution graph abstraction · p. 31

Workflow state and working memory

• Workflow state stores the user-intent summary, append-only step summaries, routing audit history, and domain-specific scientific objects passed as direct Python references.Step summaries are logged as execution traces, while heavyweight scientific objects remain outside LLM prompts.Source: Workflow state and working memory · p. 31

Typed routing configurations

• Routing configurations validate each successor name and its input parameters, rejecting missing, unknown, or out-of-range values before computation begins.Each configuration also records reasoning and confidence, with confidence constrained to [0, 1].Source: Typed routing configurations · p. 31

Deriving admissible transitions from type hints

• Node classes derive admissible successors from run() return-type annotations, restricting routing controllers to the recursively resolved set of declared transitions.In the frequency-calculation example, the declared successors are TDDFT, MolecularAnalysis, ImaginaryFrequencyRemoval, and PySCFGraphEnd.Source: Deriving admissible transitions from… · p. 31
• Type hints eagerly define and validate admissible successor sets, rejecting unspecified, invalid, duplicate, or mismatched transitions before graph execution.Resolved successor sets and routing schemas are cached per node class.Source: Deriving admissible transitions from… · p. 32

Schema-conditioned routing and fallback semantics

• Routing constrains the model to declared typed successors, records accepted decisions, and terminates conservatively when multiple branches cannot be selected safely.A single failed route may default to its sole successor with confidence 0.1, whereas multiple successors produce EndConfig with zero confidence; hard-exit state fields can force termination.Source: Schema-conditioned routing and… · pp. 33–34

LLM nodes with combined task–routing outputs

• LLM nodes return validated task outputs and routing decisions in one structured envelope, avoiding a second routing call while supporting streaming updates to GraphChat.These nodes declare successors through next_nodes and override graph registration for introspection and diagram generation.Source: LLM nodes with combined task–routing… · p. 34

Terminal nodes and KG persistence

• Terminal nodes build typed results, optionally persist them and their constituent knowledge-graph objects, and return workflow outputs to the calling agent.Persistence is gated by update_graph and requires the shared dependency providing the knowledge-graph endpoint.Source: Terminal nodes and KG persistence · p. 34

C GraphChat user interface

• GraphChat uses a synchronized CRDT room as the shared interface contract, enabling reactive co-observation of messages, logs, graphs, molecular structures, files, and execution state across clients.Rooms synchronize browser and backend state over WebSocket, persist full documents for reconnection, and support live read-only sharing, collaboration, forking, and multi-device access.Source: C GraphChat user interface · pp. 35–37

D.1 LLM model specifications

• Table S1 records per-million-token pricing and maximum context windows for the benchmarked models, excluding cached-token pricing.Source: Table S1 · p. 38
• Model-specific configurations included provider, temperature, and reasoning settings such as low reasoning effort for gpt-5, gpt-5.1, and gpt-5.2.Source: §D.1 · p. 38

D.2 User prompts

• The benchmark used two difficulty levels spanning carbocation energetics, ring strain, pKa prediction, TDDFT, and organic and inorganic compound analysis.Source: §D.2 · pp. 40–41

D.3 Rubrics for LLM judge

• The evaluation rubric checked task-specific correctness, including structures, convergence, energies, frequencies, reference choices, pKa values, and reported outputs.Source: §D.3 · p. 43

D.4 Pass k analysis

• The analysis defines pass@k as the probability that at least one of k runs passes and passˆk as the probability that all k runs pass.Source: §D.4 · p. 44
• gpt-5 achieved pass@3 of 0.99 and passˆ3 of 0.54 under the joint numerical-and-rubric criterion.Source: §D.4 · p. 44
• A run passed only when its numerical score was 1.00 and its rubric-based score exceeded 0.90.Source: §D.4 · p. 44

D.5 Statistics plots

• The benchmark’s sequential quantum-chemistry tasks favored execution graphs that retained high task scores while using fewer requests, tokens, and wall-clock time than the earlier multi-agent architecture.Source: §D.5 · p. 45
• Additional metrics measure final-call context-window saturation, error-recovery cost, and carryover tokens across agentic traces.Source: §D.5 · p. 45
• The cost calculations use raw token counts and original provider pricing rather than cached-token rates, although additional caching or local models could reduce cost.Source: §D.5 · p. 45
• GPT models generally batch concurrent tool calls, whereas Claude models more often interleave tool calls with incremental reasoning, making tool-model alignment important for efficiency.Source: §D.5 · p. 45
• Figures S5–S8 partition tokens and costs by generation source or operational role, including prompts, tool results, reasoning, calls, retries, and carryover.Source: Figure S5 · p. 47, Figure S6 · p. 48, Figure S7 · p. 49
• Figures S9–S12 show per-model distributions of request counts, final-call tokens, context pressure, and reasoning tokens across tasks and levels.One minimax-m2 Ring Strain Level 2 run used 102.91k reasoning tokens and was omitted because of the plot’s y-axis limit.Source: Figure S9 · p. 51, Figure S10 · p. 51, Figure S11 · p. 52
• Figures S13 and S14 compare actual tool calls with corresponding LLM requests by model, task, level, and tool type.Source: Figure S13 · p. 53, Figure S14 · p. 53

D.6.1 Summary of results

• The bare-agent inorganic workflow optimized SMILES-derived molecules and performed tighter single-point calculations, but misidentified symmetry, mishandled Löwdin analysis, omitted frequency checks, and produced an incorrect chlorine-trifluoride geometry.Source: §D.6.1 · p. 54
• The bare-agent pKa workflow produced -5.0, outside the rubric’s correct range of -2.70 to 1.50, after using an inaccurate COSMO-based solvation protocol.Source: §D.6.1 · p. 54
• Each exploratory comparison used one gpt-5 trace, limiting the results to recorded examples rather than performance distributions.Source: §D.6.1 · p. 54
• The inorganic and pKa bare-agent runs took approximately 40 minutes and 16 minutes, versus about 3 minutes and 5 minutes for corresponding Gráfico runs.Their trace-token counts were approximately 650k and 450k, compared with 25k and 122k for Gráfico.Source: §D.6.1 · p. 54
• The resulting generated functions suggest a route for bootstrapping new tools, but their scientific omissions require validation before reuse.Source: §D.6.1 · p. 54

E Use case extension 1: Boltzmann-weighted spectroscopic properties

• The runtime combined CREST conformer discovery, QCG solvation, DFT refinement, TDDFT, parallel branches and typed-state transfer for ensemble spectral analysis.ConceptualAtoms identifiers enabled reproducible handoff without repeatedly serializing large coordinate payloads through the LLM context.Source: E Use case extension 1:… · p. 56

E.1 Boltzmann-weighted spectroscopic properties in implicit solutions

• 412.8 nm in water and 429.9 nm in n-heptane were the ensemble absorption maxima, so conformational averaging only weakly modified the 17 nm solvent-dependent shift.The water ensemble was distributed over four conformers, whereas one conformer accounted for 95.2% of the n-heptane ensemble.Source: E Use case extension 1:… · p. 56

E.2 Boltzmann-weighted spectroscopic properties with explicit solvation

• Explicit water with SMD shifted the lowest bright absorption from 9.47 eV (131.0 nm) to 9.21 eV (134.7 nm), while the ensemble maximum shifted from 121.8 to 130.9 nm.Two microsolvated structures contributed 63.4 and 36.6% of the ensemble, and direct solute–solvent interactions were consistent with the spectral changes.Source: E Use case extension 1:… · p. 56

F.1 MOF execution graph node implementations

• The MOF execution graph routes typed state through CIF, knowledge-graph, and construction stages, with intermediate and final artifacts optionally persisted through an OGM layer.Routing is selected by an LLM API from graph state and summarized user intent, so unexpected behavior may occur.Source: §F.1 · p. 58

F.1.1 CIF decomposition to OGM building blocks

• Experimental CIFs are classified by topology, decomposed into building blocks, deduplicated structurally, and restricted to exactly one unique metal node and organic linker.Classification accepts only single-label, catenation = 1 structures; deduplication combines PORMAKE hashing with StructureMatcher verification.Source: §F.1.1 · p. 58

F.1.2 OGM instances and ConceptualAtoms

• ConceptualAtoms separates real atoms from docking points, supports ASE and QCElemental round-trips, and preserves object identity across serialization through instance_iri lookup.The same identity can persist between in-memory Python state and the external knowledge graph, including through ASE metadata.Source: §F.1.2 · p. 59

F.1.3 MOF combinatorial search algorithms

• The combinatorial-search algorithms propose new MOFs either by recombining components proven on a topology or by matching equivalent local-structure roles across topologies.Both approaches require distinct metal and linker roles, exclude already constructed combinations, and support optional VALUES constraints.Source: §F.1.3 · pp. 59–62

F.1.4 Connecting combinatorial search to PORMAKE construction

• CombinatorialSearch resolves SPARQL-returned component IRIs into fully hydrated OGM objects and queues them for PORMAKE construction.Recursive knowledge-graph materialization retrieves topology, metal-node, and linker objects with available nested geometry and motifs.Source: §F.1.4 · p. 62
• Combinatorial search canonicalizes repeated entities so shared IRIs map to the same Python objects, then queues topologies and building blocks for downstream PORMAKE construction.The queue contains topologies, node and edge building blocks, and names for BuildMOFBasic.Source: §F.1.4 · p. 63

F.1.5 Direct construction from PORMAKE building-block Identifiers

• Direct PORMAKE identifiers bypass CIF decomposition, resolve to topology and building-block OGM objects, and enter the same canonicalized build queue as knowledge-graph-driven inputs.Uniform downstream handling supports BuildMOFBasic, MLFFGeomOpt, and ZeoppAnalysis regardless of entry point, with routing controlled by user intent and graph state.Source: §F.1.5 · p. 63

F.1.6 MOF construction via PORMAKE

• PORMAKE assembles each queued topology–building-block triple after type assignment and RMSD validation, then wraps the periodic framework as a provenance-bearing ConstructedMOF object.Assignments exceeding PORMAKE’s configurable default RMSD threshold of 0.3 raise an error; the resulting object stores source topology, building blocks, connectivity, and CIF text.Source: §F.1.6 · p. 63

F.1.7 MLIP geometry optimization

• MLFFGeomOpt converts ConstructedMOF geometries to ASE, applies cached MLIP calculators and ASE relaxation, then updates geometry and CIF text in place while preserving instance identity.Supported calculators include MACE-MOF, Orb, MatterSim, and MACE-OMOL; optional cell relaxation and live CIF streaming are available.Source: §F.1.7 · p. 64

F.1.8 Zeo++ porosity analysis

• Zeo++ analysis records pore diameters, surface areas, pore volumes, channel dimensionality, and framework dimensionality as OGM results, or returns diagnostic error objects on failure.Analyses operate on each ConstructedMOF’s stored CIF text through the CoREMOF Python wrapper.Source: §F.1.8 · p. 64

F.1.9 Knowledge graph persistence

• Stable IRIs and interconnected OGM objects let completed MOF structures, components, geometries, and Zeo++ results persist as RDF for later SPARQL queries and combinatorial searches.The MOFGraphEnd node pushes ConstructedMOF instances to the knowledge graph, preserving the full provenance chain across sessions.Source: §F.1.9 · p. 64

F.2 Chat transcripts

• When a network error interrupted one workflow, the agent isolated and reran the failed job so the complete pipeline finished without data loss.Source: §F.2 · p. 66

F.2.3 Interactive exploration

• Across five natural-language messages, the agent incrementally built and explored a MOF knowledge graph while dynamically selecting minimal workflow stages, domain tools, and graph queries.The proof of concept constrained construction to one metal node and one organic linker, while its modular architecture supports extension to more complex reticular workflows.Source: §F.2.3 · p. 66

Computational setup and structural response

• 2.12 to 2.91 Å: reduction lengthened the S–S bond, indicating that the radical anion is highly weakened and close to dissociation.All optimized structures had 0 imaginary frequencies; the calculations used B3LYP/def2-SVP with SMD implicit solvent in DMF.Source: §G.1 · p. 68

Mechanistic interpretation

• Follow-up cleavage opens an EC/ECE pathway in which PhS• can accept another electron and the reverse scan can oxidize thiolate rather than the original radical anion.The thiyl radical is described as easier to reduce than the parent disulfide, making an additional electron transfer thermodynamically accessible.Source: §G.1 · p. 69

Redox thermochemistry

• −1.788 V vs Fc/Fc+: the first cathodic event corresponds to reduction of PhSSPh to [PhSSPh]•−.The calculation was framed as a fast screen of diphenyl disulfide in DMF using a cheaper functional and basis set.Source: §G.1 · pp. 67–69
• 0.91 V: the calculated PhS•/PhS− couple is more positive than the initial disulfide reduction, placing thiolate-derived oxidation near −0.87 V versus Fc/Fc+.This difference supports a reverse wave associated with a chemically transformed species rather than the intact radical anion.Source: §G.1 · pp. 68–69

Mechanistic interpretation

• The large peak separation arises because the system is no longer oxidizing the same species that underwent the first reduction.A simple reversible process would place the anodic return near the cathodic formal potential, whereas the reduced state reorganizes substantially.Source: §G.1 · p. 69

Scope and limitations

• The model is a cheap screening calculation: absolute potentials may carry uncertainty of a few tenths of a volt, and explicit electrolyte or counterion effects are omitted.The continuum-only treatment may underestimate stabilization of thiolate and the favorability of cleavage.Source: §G.1 · pp. 68–69

G.2 Luca-Turin smell theory

• Ethanol and ethanethiol have markedly different vibrational fingerprints, especially in O–H versus S–H and C–O versus C–S modes.The computed O–H stretch is 3879 cm^-1 for ethanol, whereas the S–H stretch is 2764 cm^-1 for ethanethiol; C–S modes also occur lower than C–O modes.Source: §G.2 · p. 71
• Under Turin’s framework, these shifted vibrational bands would produce distinct receptor signatures for ethanol and ethanethiol.Source: §G.2 · p. 71
• Replacing oxygen with sulfur shifts characteristic vibrations to lower frequencies because sulfur is heavier and the corresponding bonds differ in stiffness.Source: §G.2 · p. 71
• The spectral contrast offers a vibrational rationale for ethanol’s mild alcoholic odor and ethanethiol’s intense sulfurous or skunky odor.Source: §G.2 · p. 71
• The argument is not definitive proof of vibrational olfaction, because receptor shape, electrostatics, binding interactions and receptor-specific chemistry remain competing considerations.Source: §G.2 · p. 71

H.1.1 Task specification and trace scope

• The requested graph had to extract MOF synthesis data, retrieve and prepare structures, calculate Zeo++ properties, compare them with reported BET values, and produce tables, plots and discrepancy notes.Source: §H.1.1 · p. 72

H.1.2 Generated execution graphs

• The graph-construction skill imposed typed dataflow, semantic and deterministic nodes, fan-out and joins, terminal failure handling, tests and file-backed provenance.Source: §H.1.1 · p. 72

H.1.1 Task specification and trace scope

• Five papers were used for development and three were held out until each agent completed its graph and tests, with supplied CIFs used locally without live CCDC retrieval.Source: §H.1.1 · p. 72, Table S7 · p. 72
• At the stated accounting scope, Claude Code’s API-equivalent cost was 9.0-fold Gráfico’s, excluding scientific-compute, tool and infrastructure costs.Source: §H.1.1 · pp. 72–73, Table S8 · p. 73

H.1.2 Generated execution graphs

• Both workflows combined model-mediated extraction with deterministic structure preparation, P1 conversion, Zeo++ execution, numerical comparison, output generation and provenance manifests.Source: §H.1.1 · p. 72, §H.1.2 · p. 73
• Gráfico added a scientific identity gate, whereas Claude Code used more granular terminal carriers and broader test coverage; these differences were visible in code and run artefacts.Source: §H.1.2 · p. 73

H.1.3 Development-set outputs

• The development traces produced different extraction records, including three Claude Code rows lacking local deposited structures that remained explicit no_calculated_value outcomes.Source: §H.1.3 · p. 73

H.1.4 Held-out paper results and interventions

• The held-out numerical outputs are revision results after reserved-paper outcomes were observed, not a frozen-graph estimate of generalization.Source: §H.1.4 · p. 73, Table S9 · p. 74
• For aligned CAU-23 and KMF-1 structures, independently generated calculations differed by 2.1% and 0.7%, respectively.Source: §H.1.4 · p. 73
• Agreement between Zeo++ accessible surface area and BET surface area is limited because the former is static geometric output whereas the latter is fitted from adsorption data.Source: §H.1.4 · pp. 73–74

H.1.5 CIF-content mismatches as a scientific-validity gate

• On development data, Gráfico blocked both mismatched UNSW-100 and UNSW-101 paper–CIF lanes, whereas Claude Code assigned the swapped structures by accession number.Source: §H.1.5 · pp. 74–75
• On held-out data, the same gate detected reversed Zr-abtc and Zr-tptc-(Me)2 CIF identities and blocked both calculations before repair.Source: §H.1.5 · p. 75, Table S11 · p. 75
• Authorized overrides selected replacement local CIFs while preserving article identities and accession provenance, requiring multiple evidence attributes and revalidating each replacement.Source: §H.1.5 · pp. 75–76
• The identity gate cross-checked article assignments against CIF block names, formulae, atom-site composition and space groups before reporting material-specific values.Source: §H.1.5 · pp. 74–76

H.1.6 Curation, extraction variability and comparison scope

• The curation workflows differed because Gráfico preferred FSR while Claude Code preferred ASR, and both fell back to raw deposited structures when CoRE-MOF failed.Source: §H.1.6 · p. 76
• For Zr-bptc, both workflows obtained zero accessible area only after curation failure and conversion of the raw guest-containing CIF to P1.Source: §H.1.6 · p. 76
• LLM-mediated extraction varied across runs, so these single traces do not estimate extraction success; repeated authoring with frozen outputs and an audited reference set would be needed.Source: §H.1.6 · p. 76

I Prospective runtime extensions

• Future runtime extensions include asynchronous resource-aware execution, evolving semantic boundaries and long-horizon distributed networks of AI scientists.Source: Figure S16 · p. 77, Table S12 · p. 77