cs.AIFeb 19, 2026
SaveEl Agente Gráfico: A Semantic Execution Runtime for Scientific Agents
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 · Department of Computer Science, University of Toronto, 40 St George St., Toronto, ON M5S 2E4, Canada · Department of Mathematics, University of Toronto, 40 St George St., Toronto, ON M5S 2E4, Canada ·
Abstract
Large language models (LLMs) can plan scientific workflows and generate code, but these capabilities do not specify how scientific state is validated, transferred and recorded across heterogeneous computational and experimental operations. Here we present El Agente Gráfico, a semantic execution runtime for scientific agents that uses typed execution graphs to enforce admissible scientific state transitions, record provenance and confine model judgement to explicit decision points. Using the same top-level LLM and task-specific rubrics on six university-level quantum chemistry exercises, El Agente Gráfico improved performance while reducing model cost by approximately 80% and wall-clock time by more than fourfold relative to our previous multi-agent architecture. Across ensemble spectroscopy and metal-organic framework (MOF) design, the runtime supported typed-state transfer, parallel execution and cross-session persistence. Open-ended studies of a disulfide redox mechanism and the vibrational hypothesis of olfaction required the agent to define the chemical scope and computational pathway of each investigation. A graph-construction skill distilled from these studies then enabled coding agents to author typed execution graphs for MOF literature mining. Together, these results show how typed execution graphs turn reusable procedural knowledge into scientific protocols that can be inspected, revised and transferred across computational and physical systems.
Figures & tables
Figure 1: Runtime architecture of El Agente Gráfico. Users interact with a top-level agent through a natural-language interface using multi-device access with shared context (Supplementary Figs. S1 and S2 ). The agent invokes typed execution-graph as tools in which selected states are persisted to an external kg after exit. The agent can also query or update the kg as its long-term memory. Within each graph, directed edges restrict admissible transitions, and a focused llm controller resolves graph-defined routing decisions. The lower panel expands one such decision: the result from node 1 and the input schemas of admissible successors 2 and 3 form the controller context. The controller selects node 2 and returns typed input, which the runtime validates before execution; validation errors are returned for correction. Large scientific objects bypass the controller and pass between nodes in Python memory, while node outputs update the context for subsequent decisions. The GraphChat interface is shown in Supplementary Sec. C and Supplementary Video https://www.youtube.com/playlist?list=PLaUD8plXw_ecR7A1EwVAKL3pzIZqjvuVU .
Figure 2: Instantiation of the semantic execution runtime for GPU-accelerated quantum chemistry. (a) The GPU4PySCF [ 39 , 40 ] execution graph represents geometry optimization, frequency calculation, imaginary-frequency removal, molecular analysis and tddft as nodes connected by admissible transitions. Large numerical objects, illustrated by the molecular-orbital coefficient matrix, or any other wavefunction data, pass between nodes as native Python objects without entering the llm context; the controller may terminate the graph after any node once it determines that the user request has been satisfied. (b) For an aqueous pK a benchmark task, the in-graph llm controller uses the user request and preceding optimization result to select the permitted transition from geometry optimization to frequency calculation. The runtime validates the typed input before executing the selected node. The GraphChat interface reports the executed calculations and exposes the resulting files and molecular trajectory to the user.
| llm model | Numerical eval. (%) | llm Judge eval. (%) | Trace tokens | llm api ( api ) requests | Token cost (USD) | Task duration (s) | Context window saturation (%) | Error recovery cost (%) | Carryover tokens (%) |
|---|---|---|---|---|---|---|---|---|---|
| El Agente Q [ 7 ] | 88.25 (human evaluation) | 1,649,616 | 168.43 | 4.67 | 1,827 | n/a | n/a | n/a | |
| sonnet-3.7 | 93.69 | 90.94 | 284,036 | 9.14 | 0.92 | 404 | 19.46 | 5.07 | 86.30 |
| sonnet-4.5 | 96.07 | 95.67 | 320,397 | 6.58 | 1.09 | 273 | 26.79 | 9.02 | 83.28 |
| gpt-4.1 | 93.71 | 96.52 | 113,175 | 5.11 | 0.25 | 208 | 3.59 | 8.17 | 66.79 |
| gpt-5 | 98.88 | 98.50 | 83,613 | 3.32 | 0.17 | 228 | 9.78 | 9.59 | 53.21 |
| gpt-5.1 | 98.37 | 96.61 | 99,775 | 3.65 | 0.17 | 211 | 10.43 | 8.69 | 58.19 |
Table 1: Task performance and execution efficiency of llm drivers in Gráfico, averaged over 120 runs per model (six exercises, two difficulty levels, and ten repeats). The best performance is shown in bold, and the second-best performance is underlined. Human evaluators assigned the rubric-based score for El Agente Q [ 7 ] (powered by sonnet-3.7 ); an independent gpt-4o judge assigned those for Gráfico using the same task-specific rubrics. The deterministic numerical evaluator was applied only to Gráfico. Token costs exclude provider-side cache discounts. Context-window saturation, error-recovery cost and carryover tokens quantify trace-level context use; full definitions and model configurations are provided in Supplementary Secs. D.5 and D.1 . All calculations used a compute node with four H100 GPUs, with at most three calculations per GPU and four concurrent agents.
Figure 3: Typed state transfer and parallel execution for ensemble spectroscopy. Gráfico generated gas-phase and solvated conformer ensembles with crest and qcg , persisted the resulting ConceptualAtoms objects and passed their iri as lightweight handles to parallel GPU4PySCF refinement and tddft workflows. The conformer-level outputs were joined for Boltzmann reweighting and absorption-spectrum construction. The resulting plots compare merocyanine under implicit solvation and 2,3-epoxybutanol in the gas phase and an explicit water environment. Spectra were generated by Gráfico; plot titles and background grids were removed manually. Prompts are summarized; complete prompts and transcripts are provided in Supplementary Sec. E .
Figure 4: Persistent relational state for iterative MOF design. (a) The workflow acquired structures from the CoRE-MOF database by CCDC refcode [ 69 , 70 ] , decomposed them into related topology, metal-node and linker entities, proposed new combinations, constructed structures with PORMAKE [ 64 ] , optimized them with the GPU-accelerated MACE-MOF mlip [ 66 , 67 ] and calculated porosity with Zeo++ [ 68 ] . The object graph mapper preserved entity identity and relations during concurrent construction. (b) An initial session built and analysed experimental and hypothetical structures (Q1); a new session queried the accumulated state by metal node and topology without access to the preceding conversation (Q2). Prompts are summarized; complete prompts and transcripts are provided in Supplementary Sec. F .
Figure 5: Skill-guided authoring, held-out execution and diagnostic revision of an MOF-mining workflow. (a) Coding agents were equipped with a graph-construction skill and a domain curation skill. (b) The (illustrative) natural-language request to agents to develop the workflow on five papers and evaluate it unchanged on three reserved papers. (c) The Gráfico-generated graph separates LLM-mediated extraction from deterministic identity checking, structure curation, and Zeo++ calculation for comparison with reported Brunauer–Emmett–Teller (BET) values. An identity gate introduced after detecting a development-set CIF mismatch subsequently blocked an analogous held-out mismatch. Following user authorization, the revision enabled the rerun summarized in the scatter plot. Complete prompts and revision histories are reported in Supplementary Sec. H.1 .
Appendix figures & tables28 assets
Supplementary material from the paper’s appendix.
Appendix
Figure S1 : The default GraphChat interface. The chat panel (left) streams the agent’s response, reasoning, and tool-call cards; the workspace panel (right) hosts the interactive molecule viewer and the room file explorer. A live token-usage readout and a model selector are shown in the header.
Figure S2 : The default GraphChat interface accessed via phone.
| LLM model | Input tokens ($/M) | Output tokens ($/M) | Maximum context window |
|---|---|---|---|
| gpt-4.1 | 2 | 8 | 1,047,576 |
| gpt-5 | 1.25 | 10 | 400,000 |
| gpt-5.1 | 1.25 | 10 | 400,000 |
| gpt-5.2 | 1.75 | 14 | 400,000 |
| minimax-m2 | 0.2 | 1 | 205,000 |
| qwen3-max | 1.2 | 6 | 262,144 |
Table S1 : Pricing (per million tokens) and maximum context window size for benchmarked llm . Data were sourced from official provider websites for proprietary models and OpenRouter for open models. Pricing for cached tokens was not considered since cashing policies vary across providers.
| Question | Level | Weight | Task |
|---|---|---|---|
| Organic compounds | 1 | 20% | Correct input file, i.e., level of theory, required keywords, charge, and multiplicity; |
| 20% | Convergence of calculation, i.e., SCF, geometry optimization, and absence of imaginary frequency; | ||
| 20% | Report generation; | ||
| 20% | Successful extraction and documentation of all report values; | ||
| 20% | Successful generation of XYZ from SMILES. | ||
| 2 | 20% | Correct input file, i.e., level of theory, required keywords, charge, and multiplicity; |
Table S2 : The rubric used by the LLM judge when grading the benchmark questions.
| pass@k | pass ˆ k | |||||||
|---|---|---|---|---|---|---|---|---|
| LLM Model | @1 | @3 | @5 | @10 | ˆ 1 | ˆ 3 | ˆ 5 | ˆ 10 |
| gpt-4.1 | 0.81 | 0.99 | 1.00 | 1.00 | 0.81 | 0.52 | 0.34 | 0.11 |
| gpt-5 | 0.94 | 1.00 | 1.00 | 1.00 | 0.94 | 0.83 | 0.74 | 0.53 |
| gpt-5.1 | 0.87 | 1.00 | 1.00 | 1.00 | 0.87 | 0.65 | 0.48 | 0.22 |
| gpt-5.2 | 0.84 | 1.00 | 1.00 | 1.00 | 0.84 | 0.59 | 0.42 | 0.17 |
| minimax-m2 | 0.53 | 0.90 | 0.98 | 1.00 | 0.53 | 0.15 | 0.04 | 0.00 |
Table S3 : Pass@k and Pass ˆ k rates with criteria: and .
| pass@k | pass ˆ k | |||||||
|---|---|---|---|---|---|---|---|---|
| LLM Model | @1 | @3 | @5 | @10 | ˆ 1 | ˆ 3 | ˆ 5 | ˆ 10 |
| gpt-4.1 | 0.68 | 0.97 | 1.00 | 1.00 | 0.68 | 0.30 | 0.13 | 0.02 |
| gpt-5 | 0.82 | 1.00 | 1.00 | 1.00 | 0.82 | 0.56 | 0.38 | 0.13 |
| gpt-5.1 | 0.76 | 0.99 | 1.00 | 1.00 | 0.76 | 0.43 | 0.24 | 0.06 |
| gpt-5.2 | 0.82 | 1.00 | 1.00 | 1.00 | 0.82 | 0.56 | 0.38 | 0.13 |
| minimax-m2 | 0.55 | 0.91 | 0.98 | 1.00 | 0.55 | 0.16 | 0.05 | 0.00 |
Table S4 : Pass@k and Pass ˆ k rates with criteria: .
| pass@k | pass ˆ k | |||||||
|---|---|---|---|---|---|---|---|---|
| LLM Model | @1 | @3 | @5 | @10 | ˆ 1 | ˆ 3 | ˆ 5 | ˆ 10 |
| gpt-4.1 | 0.63 | 0.95 | 0.99 | 1.00 | 0.63 | 0.25 | 0.10 | 0.01 |
| gpt-5 | 0.82 | 0.99 | 1.00 | 1.00 | 0.82 | 0.54 | 0.36 | 0.12 |
| gpt-5.1 | 0.70 | 0.97 | 1.00 | 1.00 | 0.70 | 0.34 | 0.16 | 0.02 |
| gpt-5.2 | 0.75 | 0.99 | 1.00 | 1.00 | 0.75 | 0.42 | 0.23 | 0.05 |
| minimax-m2 | 0.43 | 0.81 | 0.94 | 1.00 | 0.42 | 0.07 | 0.01 | 0.00 |
Table S5 : Pass@k and Pass ˆ k rates with criteria: and .
Figure S3 : Radar plots of numerical-evaluator scores averaged over 10 runs by task and model for Level 1 and Level 2. Each polygon shows a model’s mean score across tasks.
Figure S4 : Radar plots of LLM-as-a-judge-evaluator scores averaged over 10 runs by task and model for Level 1 and Level 2. Each polygon shows a model’s mean score across tasks.
Figure S5 : Mean token composition by generation source across tasks, models, and levels. Stacked bars categorize input tokens into base prompts (system prompts, tool schemas, and user task prompts) and tool-result inputs, and output tokens into reasoning tokens, action/tool-call outputs, and text outputs.
Figure S6 : Same split as Fig. S5 , but measured in USD cost calculated from pricing provided by LLM providers.
Figure S7 : Mean token composition by role across tasks, models, and levels. Stacked bars partition total tokens into base input, normal operation (in/out), immediate retry, and carryover burden across agentic traces; rows separate Level 1 and Level 2.
Figure S8 : Same split as Fig. S7 , but measured in USD cost calculated from pricing provided by LLM providers.
Figure S9 : Distribution of LLM requests by model across tasks and levels. Boxes show per-model variability in LLM requests, with overlaid points for individual runs; facets separate tasks (columns) and levels (rows), and colours denote models.
Figure S10 : Distribution of final call tokens by model across tasks and levels. Boxes show per-model variability in final call total tokens, with overlaid points sized by number of LLM requests; facets separate tasks (columns) and levels (rows), and colours denote models.
Figure S11 : Distribution of final call context window pressure (final call tokens divided by the maximum context window) by model across tasks and levels. Boxes show per-model variability, with overlaid points representing individual runs scaled by the number of LLM requests; facets split tasks (columns) and levels (rows), with colours denoting models.
Figure S12 : Distribution of reasoning token usage by model across tasks and levels. Boxes show per-model variability in reasoning tokens (filtered to LLM models with reasoning capabilities; sonnet-3.7 has a consistent reasoning budget of 2024), with overlaid points for individual runs; facets split tasks (columns) and levels (rows), and colours denote models. One run from minimax-m2 for Ring Strain (Level 2) consumed 102.91k reasoning tokens and is omitted from the plot due to the y-axis limit.
Figure S13 : Average number of tool calls versus LLM API requests per model for the benchmark exercise. For each model, stacked bars compare three tool types ( run_pyscf_workflow , run_python_code , get_conversion_factor ), shown separately for actual tool calls and corresponding LLM requests (side-by-side groups).
Figure S14 : Tool usage versus LLM API requests by task and level for the benchmark exercise. For each task (grouped blocks) and model (ticks), stacked bars compare average tool calls (left bar) with corresponding LLM requests (right bar) for the run_pyscf_workflow , run_python_code , and get_conversion_factor tools; rows separate levels and colours encode tool type.
| Stage | Input Representation | Output Representation | Translation Mechanism |
|---|---|---|---|
| CIF → Topology | CIF file path | Topology ( ogm ) | CrystalNets.jl classification → PORMAKE topology → Topology ogm wrapper |
| CIF → Building Blocks | CIF file path | MetalNode / OrganicLinker ( ogm ) | PORMAKE MOFDecomposer → ASE Atoms fragments → deduplicate_building_blocks → create_ontomofs_building_block (splits real atoms vs docking points into ConceptualAtoms + XYZ ) |
| kg → Python Objects | SPARQL result iri | Topology , MetalNode , OrganicLinker | pull_from_kg with recursive hydration ( recursive_depth=-1 ) |
| PORMAKE IDs → ogm components | String identifiers (topology / node / linker) | Topology , MetalNode , OrganicLinker | Resolve identifiers to PORMAKE definitions → wrap as ogm objects ( ConceptualAtoms + docking points) → canonicalize and enqueue in state.build_queue |
| ogm → PORMAKE | BuildingBlock ( ogm ) | PORMAKE BuildingBlock | Lazy property access ( e.g. , cached PORMAKE BuildingBlock construction inside ogm wrapper) |
| PORMAKE → ConstructedMOF | PORMAKE Framework | ConstructedMOF ( ogm ) | from_pormake_framework → ConceptualAtoms.from_ase_atoms → provenance wiring |
Table S6 : Representation translations across the mof execution graph. Each row describes a stage in the pipeline, the input and output data representations, and the mechanism used for translation.
Figure S15 : Redox mechanism exploration for diphenyl disulfide (PhSSPh) in DMF. (i) Gráfico starts PySCF workflows to evaluate relevant charge and spin states for PhSSPh, its radical anion, thiolate, and thiyl radical species. The failed calculation for the was retried with xyz coordinates generated directly by Gráfico. (ii) Proposed electrochemical redox network explaining the large separation between cathodic and anodic peaks. Computed free energies and standard potentials versus Fc/Fc+ are shown for the key elementary steps.
| Phase | Gráfico | Claude Code |
|---|---|---|
| Development | Detected the reversed UNSW-100/UNSW-101 CIF contents and encoded composition-, name- and space-group-based identity checks. The completed graph blocked both mismatched pairs; three tests passed before reserved-paper access. | Authored the graph and completed 23 tests without an article–CIF identity check. Its final development run calculated both mismatched UNSW pairs; extraction varied across model calls. |
| First reserved-paper pass | Applied unchanged, the development-stage gate calculated CAU-23 and one KMF-1 deposition, blocked the reversed Zr pair, and blocked Zr-bptc for insufficient formula evidence; the run returned partial . | Processed the KMF-1 and Zr-framework papers without detecting the Zr reversal. CAU-23 entered a paper_failed carrier after its extraction call reached the 180-s timeout. |
| Revision after observation | Explained the conflicts and, after user authorization, added atom-site composition fallback and provenance-backed local-CIF overrides for both reversed pairs. Six tests passed, and both folders were rerun. | Increased the extraction timeout to 300 s, added one retry and reran the reserved folder. It did not add a structure-identity check or revise the local-CIF associations. |
Table S7 : Chronology of development, held-out execution and revision. The final row distinguishes results obtained with the development-stage graph from revisions made after the reserved-paper outcomes had been observed.
| Measure | Gráfico | Claude Code |
|---|---|---|
| Scope | Three consecutive traces covering initial authoring and reserved-paper execution, explanation of the blocked identities and the user-authorized repair. | Session start through the workflow completion. |
| Model requests | 42 calls to gpt-5.6-sol . | 154 billable model iterations: 27 on Claude Fable 5 and 127 on Claude Opus 4.8, including fallback iterations. |
| Input-side accounting | 3,234,165 input tokens, including 3,111,841 cache-read tokens (96.2%) and 122,324 uncached tokens. | 25,623,337 token events: 9,675 base-input tokens, 748,189 one-hour cache-write tokens and 24,865,473 cache-read tokens (97.0%). |
| Output tokens | 32,141, including 7,985 reasoning tokens. | 139,253. |
| API-equivalent model cost | US3.131771), summed from public-price cost fields in the trace telemetry. | US28.104450), reconstructed from recorded usage and public API rates. |
Table S8 : Trace-level model usage and API cost. Counts are summed over requests and therefore represent accounting events rather than unique text. For the Claude Code input-side total, base input, one-hour cache writes and cache reads are separate billable categories. Reasoning tokens are a subset of the Gráfico output count.
| Dimension | Gráfico | Claude Code |
|---|---|---|
| Graph topology | Separate paper-parsing and paper-extraction maps, each followed by a join; a subsequent per-record identity, preparation and calculation map terminated in a result join. | A paper-level map combined parsing and extraction before a paper join; a second map processed each (MOF, CCDC) pair before a result join. |
| Failure representation | Invalid identities, preparation failures and calculation failures produced result rows. Extraction failure was retained in run state and provenance but did not produce a paper-level result carrier. | Explicit carriers represented analysed, paper-failed, no-CCDC and CIF-not-found lanes; preparation and calculation exceptions were also retained. |
| Structure identity | Compared deposition number, formula-derived diagnostic elements, recognizable block name and space group. A later revision added atom-site composition fallback and explicit local-CIF overrides that were revalidated. | Resolved each mined identifier to the matching local filename. It did not compare the paper-derived framework identity with CIF formula, name or space group. |
| Solvent removal | Selected FSR when CoRE-MOF outputs and checks passed; otherwise retained the raw CIF with a warning. ASR was not selected automatically. | Selected ASR whenever available, then FSR, and finally the raw CIF if cleaning failed. |
| Zeo++ settings | High-accuracy mode; Å probe radius; 4,000 surface-area and pore-volume samples per atom. | High-accuracy mode; Å probe radius; 2,000 surface-area samples and 50,000 pore-volume samples per atom. |
| Tests | Three tests before the first reserved-paper run, a fourth dry-run test after that pass, and six after the identity-repair revision. | Twenty-three unit and integration tests before the final reserved-paper rerun. |
Table S9 : Implementation choices in the two generated MOF-mining workflows. Test counts refer to the final state of each single authoring trace and are not a standardized software benchmark.
| Framework | Reported BET | Gráfico | Claude Code |
|---|---|---|---|
| PRI-1 | 317.3 (CO 2 ) | 0; CCDC 641709, FSR | 0; CCDC 641709, ASR |
| ALF | 588 | 0; CCDC 2098663, FSR | 0; CCDC 2098663, ASR |
| ZJU-300 | 796.2 | 0; CCDC 2218592, FSR | 0 for CCDC 2218592–2218594, ASR; the final extraction omitted the BET value |
| Ca-tcpb | 260.65 | 107.59; CCDC 2277051, FSR | 102.17; CCDC 2277051, ASR |
| UNSW-100 | 1,490 | Blocked at article-reported CCDC 2448651 by identity checks | 0; local 2448651.cif, raw; its contents mismatch article-defined UNSW-100 |
| UNSW-101 | Not numerically reported | Blocked at article-reported CCDC 2448652 by identity checks | 933.82; local 2448652.cif, ASR; its contents mismatch article-defined UNSW-101 |
Table S10 : Development-stage surface-area outputs. Values are reported BET and calculated Zeo++ accessible surface area in . These runs preceded access to the reserved paper contents.
| Framework | Reported BET | Gráfico post-diagnostic | Claude Code final | Qualification |
|---|---|---|---|---|
| CAU-23 | 1,250 | 1,367.69; CCDC 1878820, FSR | 1,396.96; CCDC 1878820, ASR | Aligned identity; and relative to the reported BET. Claude’s value followed the timeout/retry revision. |
| KMF-1 | 1,130 | 1,033.80; CCDC 1984701, FSR | 1,040.98 and 1,014.92; CCDC 1984701 and 1984702, ASR | Aligned identities; , and , respectively. |
| Zr-bptc | 1,030 | 0; CCDC 1557153, raw | 0; CCDC 1557153, raw | Both curation paths failed and retained the solvated deposited structure; zero accessibility is not a comparison with the activated experimental material. |
| Zr-abtc | 1,318 | 1,191.97; article CCDC 1557154, local 1557155.cif, FSR | 2,907.24; local 1557154.cif, ASR | The supplied 1557154 CIF matches article-defined Zr-tptc-(Me) 2 rather than Zr-abtc; the Claude Code value therefore does not evaluate the reported Zr-abtc material. |
| Zr-tptc-(Me) 2 | Not reported | 2,566.82; article CCDC 1557155, local 1557154.cif, FSR | 1,496.95; local 1557155.cif, ASR | The supplied 1557155 CIF matches article-defined Zr-abtc rather than Zr-tptc-(Me) 2 ; no experimental BET was mined for the latter. |
Table S11 : Results for the three held-out papers. Surface areas are in . Raw denotes a deposited structure for which CoRE-MOF curation failed. Values calculated from local CIF contents that conflict with the article-reported assignments are retained to show the downstream consequence but are not treated as aligned cross-agent comparisons.
| Dimension | Evidence in this study | Open requirement | Future evaluation |
|---|---|---|---|
| Protocol semantics | Typed state, admissible transitions and explicit failure carriers | Static and domain-specific conformance checks, with versioned schema migration | Held-out protocol suites and regression testing across workflow revisions |
| Protocol authoring | Skill-guided graph generation by two coding agents | Evidence-gated revision of graphs and skills | Repeated cross-model authoring with independent scientific audit |
| Execution substrate | Native-object transfer, persistent relational state and parallel graph lanes | Isolated, resource-aware and distributed execution that preserves identity and provenance | Cross-environment replay, recovery and resource accounting |
| Physical grounding | Digital-twin-to-robot protocol transfer and conversion of laboratory feedback into a tested constraint | Typed multimodal feedback, calibration and uncertainty-aware twin updates | Repeated closed-loop experiments on held-out targets |
Table S12: Current evidence and open requirements for scientific execution runtimes. Each row separates a capability examined in this study from an unresolved requirement and a future evaluation that could test it. The ordering does not imply a maturity sequence.
Figure S16 : Current stage and roadmap of future work showing: (1) structured execution of typed execution graphs and tools (as implemented in the current Gráfico system); (2) asynchronous and resource-aware execution to enable proactive agents; (3) semantic boundary evolution allowing rapid extension of tools and ontologies; (4) long-horizon agents towards a distributed network of AI scientists.