CASCADE: Cumulative Agentic Skill Creation through Autonomous Development and Evolution

1 Introduction

• The framework targets a gap in scalability because current pipelines require human-curated tool catalogs and task-specific prompts, while autonomous tool generation remains narrow in scope.CASCADE’s predefined tools are domain-agnostic, supporting potential transfer to fields such as software engineering and biology.Source: §1 · pp. 2–3
• Existing scientific agents mainly support literature search, hypothesis generation, and data analysis, while complex long-horizon computational and physical experiments remain challenging.Current systems typically depend on predefined tools or early-stage tool-generation approaches.Source: §1 · p. 2
• CASCADE introduces an early transition from “LLM + tool use” to “LLM + skill acquisition” for AI co-scientists.The framework facilitates cumulative executable skills rather than relying solely on problem-specific tools and prompts.Source: §1 · pp. 2–3
• CASCADE cultivates continuous learning and self-reflection as meta-skills through general problem-solving methods, natural-language interaction, and multi-turn collaboration.It also uses session-wise and consolidated memory to retain conversations, information, experience, and skill sets.Source: §1 · p. 3
• CASCADE is evaluated with SciSkillBench and real-world research scenarios spanning materials analysis, autonomous laboratory work, and selective reproduction of published results.The benchmark targets materials science and chemistry research, while the demonstrations include piezoelectricity, MLIP prediction differences, materials synthesis and characterization, and Li intercalation voltages.Source: §1 · p. 3

2.1 CASCADE architecture

• DeepSolver uses web search, code extraction, dependency management, conditional parallel debugging, output selection, memory, and introspective tools to acquire and refine skills.The Orchestrator routes straightforward tasks to SimpleSolver and complex tasks to DeepSolver, while errors in SimpleSolver can trigger escalation.Source: §2.1 · pp. 3–4

2.2 Evaluating DeepSolver on scientific tasks

• 116 SciSkillBench tasks span 76 data-oriented and 40 computation-oriented problems, testing tool discovery, chaining, undocumented queries, and difficult scientific reasoning.The benchmark includes 58 Level 0 tasks with procedural guidance and 58 Level 1 tasks requiring greater autonomy.Source: §2.2 · p. 4, CASCADE · p. 6

CASCADE

• DeepSolver improves overall success over Search&Debug by 17.53% on average and reaches 89.66% Pass@3 with GPT-4.1-mini versus 87.93% for Claude Code.The GPT-4.1-mini DeepSolver run also completes in 190 s versus 239 s for Claude Code, despite lower input and output costs.Source: Table 1 · p. 7, CASCADE · p. 8
• DeepSolver achieves the highest accuracy across all six task categories and declines less with increasing difficulty than Native and Search&Debug.For example, Qwen3-Coder-30B with Search&Debug has zero pass rate on difficult tasks, whereas DeepSolver achieves nearly 50%.Source: CASCADE · p. 8, Figure 2 · p. 9

Automated scientific reasoning: piezoelectricity determination

• CASCADE independently determines piezoelectricity and predicts systematic MLIP errors by selecting external materials tools, formulating hypotheses, executing experiments, and analyzing results.The piezoelectricity workflow identifies space group I4/mmm, point group 4/mmm, and centrosymmetry, while the MLIP task concerns GGA/GGA+U and r2SCAN CHGNet predictions.Source: Automated scientific reasoning:… · p. 10, Hypothesis-driven discovery:… · p. 10, Figure 3 · p. 11

Laboratory automation: autonomous synthesis and characterization

• In the autonomous lab, CASCADE learns undocumented Alab-GPSS functionality, implements a fallback for stoichiometric balancing, and submits the Li2Fe0.8Ni0.2Cl4 synthesis workflow.For ionic-conductivity fitting, its algorithm is approximately 60-fold faster and achieves R2 0.991 versus 0.989 for expert-written code.Source: Laboratory automation: autonomous… · pp. 11–13, Figure 4 · p. 12

Collaborative research with memory: reproducing published battery calculations

• Human guidance and memory let CASCADE transfer experience across conversations, extend LixCoO2 voltage calculations to four SevenNet models, and reproduce results from the referenced article.The system learns structure-generation procedures, retrieves them later, computes the additional 0 < x < 0.5 range, and generates comparative plots.Source: Collaborative research with memory:… · p. 13, Figure 5 · p. 14

• CASCADE cultivates and accumulates executable scientific skills through continuous learning and self-reflection, advancing beyond static “LLM + tool use.”The framework also uses session-wise and consolidated memory to support evolution across interactions.Source: §3 · p. 15
• DeepSolver’s real-world demonstrations traverse literature search, hypothesis generation, experimental execution, and data analysis with customizable computational and physical experiments.These proof-of-concept applications include autonomous research workflows rather than isolated tool calls.Source: §3 · pp. 15–16
• The workflow still falls short of autonomous end-to-end discovery with truly novel insights, requiring iterative trial, error, and optimization as experience and data accumulate.The authors frame autonomous discovery as a future goal rather than an achieved capability.Source: §3 · p. 15
• Long-horizon discovery requires flexible, robust workflows, improved human-agent collaboration and memory, and better alignment between documented methods and practical materials-research execution.Tool selection and sequencing remain difficult, while limited practical data on scientists’ simulation and experimental reasoning constrains reproduction and new research.Source: §3 · pp. 15–16
• CASCADE’s skills can be shared as learnable external tools across agents and scientists, supporting accumulation, transmission, and redeployment of scientific knowledge.The authors connect this sharing model to larger-scale human-agent cooperation and collective scientific discovery.Source: §3 · pp. 15–16

4.1.1 MCP servers and tools

• CASCADE provides agents with web search, code research, introspection, execution, workspace management, and memory tools through specialized MCP servers.The system exposes 16 problem-solving tools across four server capabilities, with direct function implementations also available.Source: §4.1.1 · p. 16

4.1.2 Conversational system with memory and tracing

• The Orchestrator retrieves memories, clarifies incomplete queries, and routes straightforward problems to SimpleSolver or complex problems to DeepSolver.Session state, user-scoped memory, feedback, and MLflow tracing support personalized interactions and workflow observability.Source: §4.1.2 · pp. 16–17

4.1.3 SimpleSolver

• SimpleSolver adapts remembered or known solutions, verifies and installs packages, executes code, and delegates failures to DeepSolver.It is a pathway within the Orchestrator rather than a separate agent architecture.Source: §4.1.3 · p. 17

4.1.4 DeepSolver

• DeepSolver researches solutions, executes code once, launches parallel debugging when needed, and selects and formats the best resulting answer.Its workflow combines web research, code extraction, introspection, package management, knowledge-graph exploration, local inspection, diagnostics, and result processing.Source: §4.1.4 · pp. 17–19

4.2.1 Benchmark tasks

• SciSkillBench contains 116 materials-science and chemistry tasks spanning 76 data-oriented and 40 computation-oriented problems across six categories.The evaluation uses three repetitions across 46 configurations, isolated subprocesses, tolerance-based answer checking, and pass@3 difficulty stratification into 37 simple and 21 difficult questions.Source: §4.2.1 · p. 19, §4.2.3 · pp. 19–20, §4.2.4 · p. 20

4.2.2 Baselines

• The evaluation compares Native, Search&Debug, Claude Code, CASCADE, and two ablations under independently repeated, isolated benchmark conditions.Native lacks tools, Search&Debug adds web search and iterative debugging, and Claude Code runs independently in reset Docker containers.Source: §4.2.2 · p. 19, §4.2.3 · pp. 19–20

Data availability. / Code availability.

• SciSkillBench and CASCADE are publicly available through Figshare and GitHub.The methods section provides links for both the benchmark suite and source code.Source: Data availability · p. 20, Code availability · p. 20

A.1 MCP server infrastructure and tools

• CASCADE uses Tavily search to discover documentation, code examples, and implementation resources for agent problem solving.The search capability is provided through the Tavily MCP server.Source: §A.1 · p. 23
• Its memory server combines vector semantic search with Neo4j entity relationships while preserving original content verbatim in the vector store.The save_to_memory and search_memory tools support memory updates, deletion, semantic retrieval, and graph-based relationship retrieval.Source: §A.1 · p. 23
• The research server extracts code from diverse documentation sources and retrieves relevant code blocks through embedding-based similarity search.It also supports contextual extraction, optional summaries, introspection, runtime probing, and knowledge-graph construction and querying.Source: §A.1 · p. 23
• CASCADE analyzes packages statically and at runtime, using introspection, probes, AST-derived knowledge graphs, and Cypher queries to diagnose and explore code.These tools expose package, class, method, function, attribute, import, hierarchy, and signature information.Source: §A.1 · pp. 23–24
• The workspace server manages conda, venv, and uv environments while restricting operations to the project root and blocking benchmark directories to prevent solution leakage.Package-management tools inspect installed versions, installation paths, and module locations, and install missing dependencies.Source: §A.1 · p. 24
• CASCADE provides execution and file-management tools for Python code, shell commands, scripts, and text files, although save_file is not used.Execution captures standard output and errors in temporary files within the code storage directory.Source: §A.1 · p. 24

A.2 Underlying principles of the architecture design

• DeepSolver represents scientists’ deep thinking and experimentation, while human-agent collaboration and memory consolidation represent discussion and learned-skill retention.The architecture is explicitly designed around acquiring, refining, and remembering executable problem-solving skills.Source: §A.2 · p. 24
• DeepSolver’s research stage parallels scientists’ online searches, code-example selection, and verification of imports, classes, methods, and functions.The quick_introspect tool is intended to reduce hallucinated software interfaces.Source: §A.2 · pp. 24–25
• The execution and debugging stages mirror hands-on trials followed by local fixes, broader code exploration, knowledge-graph inspection, package inspection, and continued research.Afterward, the Output Processor summarizes information for further consolidation.Source: §A.2 · p. 25
• Context engineering supplies reminders about tool invocation and suggests subsequent tools when agents fail to use an appropriate capability.One example recommends query_knowledge_graph after parse_local_package.Source: §A.2 · p. 25

B.1.1 SciSkillBench examples

• Each SciSkillBench task stores queries, sources, input and output specifications, answers, tolerances, units, solution processes, references, and notes.Level 0 queries provide some function names, whereas level 1 queries require more independent solution planning.Source: §B.1.1 · p. 25
• The benchmark includes both guided and concise versions of scientific queries, allowing comparison between instruction-rich and independently solved task formulations.The examples include paired level 0 and level 1 prompts for thermodynamics, formation-energy, and MLIP tasks.Source: §B.1.1 · pp. 25–27
• An MLIP example requires downloading a pretrained MACE model and sample data, loading the model, running batched aspirin inference with batch size 8, and returning energy and forces for the first structure.The task uses the InstaDeepAI mlip library and specified HuggingFace and archive paths.Source: §B.1.1 · p. 26
• A thermodynamics example asks agents to construct a phase diagram and identify the Na2O–TiO2 product phase with the most negative reaction delta G at 600K.The required output is an ordered list containing the product composition and delta G rounded to four decimal places.Source: §B.1.1 · p. 26
• A CHGNet example requires computing Ti3AlC2 formation-energy error after structural relaxation against a Materials Project R2SCAN value.The workflow requires retrieving constituent structures, relaxing them with the pretrained R2SCAN CHGNet model, and accounting for atom counts when computing formation energy per atom.Source: §B.1.1 · p. 27

• Other examples require loading the OBELiX dataset and transforming a crystal structure before comparing space-group numbers.Source: §3 · p. 28
• Out-of-sync pymatgen-db and Materials Project documentation may confuse agents during package-based tasks.Source: §3 · p. 28
• Example tasks require querying MongoDB-backed VASP calculations and computing the final structure’s volume per atom.Source: §3 · p. 29

• Additional tasks involve downloading published-paper datasets for later processing and analysis.Source: §B.1.2 · p. 29
• These tasks were excluded from the main SciSkillBench evaluation because dataset URLs produced practical access problems, including “Too Many Requests” errors.Source: §B.1.2 · p. 29
• DeepSolver with OpenAI O3 was nevertheless verified to complete the additional tasks successfully.Source: §B.1.2 · p. 29

• The P value measures the proportion of correct responses across models, with values at least 0.5 classified as simple and lower values as difficult.Source: §B.1.3 · p. 30
• Examples include molecular-property calculations using ORCA, xTB, ASE, and RDKit, covering charges, electrophilicity, energy scans, atomization, adsorption, and fragments.Source: §B.1.3 · p. 31
• Materials-focused examples require pymatgen and related tools for decomposition analysis, elemental fractions, and crustal-abundance calculations.Source: §B.1.3 · pp. 30–31

• DeepSolver integrates continuous learning and self-reflection, while the ablations remove the corresponding tools and prompts to isolate each meta-skill.Source: §B.2 · p. 31
• Across models, self-reflection often contributes more than continuous learning, although weaker coding models can benefit more from continuous learning.Source: §B.2 · p. 31
• DeepSolver consistently achieves the best overall performance, and for more than half the models it is faster and more accurate than the no-continuous-learning configuration.Source: §B.2 · pp. 31–32
• Among nine evaluated models, six achieve higher overall success with the no-continuous-learning configuration than with the Search&Debug baseline.Source: §B.2 · p. 32

• Demonstration files include task queries, reference solution code, agent-generated answers, and associated figures.Source: §B.3 · p. 32
• Table B1 compares DeepSolver with No Self-Reflection and No Continuous Learning configurations.The comparison covers all questions and separate 0-Level and 1-Level subsets.Source: Table B1 · p. 33
• The ablation reports Success Rate, Pass@1, Pass@2, Pass@3, and average completion time.Bold values identify the best performance across the three configurations.Source: Table B1 · p. 33