Metal-organic frameworks (MOFs) and covalent organic frameworks (COFs) are important functional materials for energy, environmental, and sustainability-related applications^[1,2]. Their ordered pores, high surface areas, and tunable structures make them attractive for applications such as gas separation^[3], catalysis^[4], sensing^[5], and resource recovery^[1].

Unlike many conventional porous solids, MOFs and COFs can be assembled from molecular or coordination building units^, allowing their pore size, shape, functionality, and host-guest interactions to be tailored with precision^[6,7]. This programmability, however, also creates a highly complex discovery space governed by coupled structural, synthetic, and performance-related variables^[1,8]. MOF and COF discovery is a multifactorial challenge that requires balancing synthesis, crystallinity, porosity, stability, and performance, all of which are tightly coupled and sensitive to building-unit selection and experimental protocols^[1,9].

To cope with the growing complexity of crystalline porous-materials research, researchers have increasingly relied on data-driven approaches, particularly machine learning and high-throughput computation^[8,9]. These methods have accelerated property prediction and virtual screening, but they generally address individual steps of the discovery workflow rather than the workflow as an integrated process. Most models map a fixed representation to a predefined property and require a curated dataset prepared in advance. They generally do not determine what information is missing, choose which computational or experimental tool to use, reconcile contradictory evidence, revise a failed strategy, or coordinate simulation with experimentation^[10-12].

Artificial intelligence (AI) agents offer a different paradigm. They are goal-directed systems that can reason about tasks, use tools, maintain context, and learn from feedback^[10]. Although an large language model (LLM) can provide the reasoning and communication layer, a scientific agent also requires connections to external resources, such as databases, predictive models, simulators, robotic platforms, and characterization instruments. Thus, an LLM alone is not necessarily an AI agent; what defines an agent is its ability to act and adapt within a workflow^[10,12].

In this review, we use the term agent in a practical sense to describe goal-directed AI systems that can interpret scientific tasks, use relevant information or tools, act within a workflow, and adjust subsequent steps based on intermediate results. To organize this review, we classify AI agents for MOFs and COFs along two practical dimensions: scientific grounding and experimental agency. Scientific grounding refers to the extent to which an agent’s decisions are informed and constrained by external evidence, such as literature, databases, knowledge graphs, simulations, or domain-specific tools. Experimental agency refers to its participation in experimental execution and its use of experimental outcomes to guide subsequent actions. Based on these dimensions, we divide the field into three stages, as shown in Figure 1.

Stage 1 comprises LLM-native, human-mediated systems in which researchers select external resources, coordinate specialized tools, and carry out experiments. Stage 2 comprises database-grounded, tool-using agents that can access external knowledge and coordinate digital resources and computational workflows, while physical experiments remain human-mediated and outside the automated decision loop. Stage 3 comprises experiment-integrated, feedback-driven platforms in which agent-generated experimental instructions are executed through robotic or laboratory automation systems, and experimental outcomes are returned directly to the decision-making system to guide subsequent actions. The transition from Stage 1 to Stage 2 primarily reflects greater scientific grounding through access to external data and digital tools, whereas the transition from Stage 2 to Stage 3 primarily reflects greater experimental agency through the integration of automated experimental execution and feedback into the decision loop. Representative examples of each stage are shown in Figure 1 and discussed in the following sections.

Scientific AI agents are better viewed as orchestration layers than as standalone chatbots. Relevant studies in adjacent materials domains further illustrate the architectural requirements for scientific agents in porous-materials research. In practice, they connect databases, predictive models, simulators, experimental equipment, and human experts^{[11,12,21,22]}. Their capabilities can still be grouped under perception, reasoning, action, and learning, but the more useful question in science is whether they can obtain dependable information, choose suitable tools, retain relevant context and update decisions from the evidence those tools return.

The first requirement is therefore a reliable information foundation. Databases sit at the center of this architecture for materials research. They need to connect material identity, experimental conditions, structures, properties, metadata, and provenance so that results from different studies can actually be compared^[23-25]. For porous materials, that means linking crystal structures, synthesis and activation protocols, sample history, stability, performance, computed descriptors, uncertainty, and provenance to persistent material identities.

A reliable information foundation alone, however, is not sufficient. Reliable agents also need scientific tools, multimodal perception, and action traces that can be audited. Thermodynamic, electrochemical, density functional theory (DFT), and microkinetic calculations can check or narrow language-model proposals^[21,26]. Many important records in materials science appear in figures, diffraction and spectroscopic data, tabulated metadata, and provenance-rich experimental records rather than in prose alone^[27,28]. Reproducibility also depends on clear records of data versions, prompts, tool calls, failures, and human intervention^[29]. Together, these data, tool, perception, and provenance layers provide the technical foundation on which reliable MOF and COF agents can be built.

Across these layers, agent memory provides the mechanism by which information, tool outputs, and experimental states are retained and reused across successive reasoning and action steps. Here, memory refers to information maintained within the active context or stored externally for subsequent retrieval and use^[30]. Stage 1 systems have limited explicit memory and rely mainly on the model’s parametric knowledge, the immediate context window, and information supplied by researchers. Such information is often transient, difficult to update, and not readily auditable. In Stage 2, memory becomes increasingly externalized through machine-actionable resources, including databases, literature-derived records, knowledge graphs, and retained tool outputs that can be retrieved, corrected, and reused across tasks^[31]. In Stage 3, this external memory further incorporates machine-readable experimental states, action histories, and measured outcomes, allowing prior experiments to directly inform subsequent decisions^[32]. This progression reflects increasing scientific grounding from Stage 1 to Stage 2 and increasing experimental agency from Stage 2 to Stage 3.

In this review, Stage 1 denotes LLM-native, human-mediated systems in which the model supports task decomposition, literature interpretation, data organization, and planning. Tool and experimental actions remain human-defined and human-supervised, although automation may be used. Human feedback may revise later prompts, but it does not constitute an autonomous, machine-readable loop that selects the next experimental action.

Zheng et al. developed the GPT-4 Reticular Chemist as a three-phase LLM-guided workflow for reticular chemistry^[13]. GPT-4 generated project blueprints and experimental steps, and a researcher returned successful and failed outcomes to later prompts. The workflow guided the discovery of an isoreticular MOF series, showing how an LLM can support project planning and iterative experimental decisions under human supervision^[13] as shown in Figure 2.

The same group later built the ChatGPT Research Group around role-specific LLM assistants, Bayesian optimization, and researcher-coordinated robotic operation^[14]. The workflow iteratively proposed and evaluated crystallization conditions for MOFs and COFs. This study extends LLM-assisted planning to a more structured optimization workflow while retaining researcher oversight. In this workflow, experimental crystallinity measurements, including powder X-ray diffraction (PXRD) analysis, were communicated through researcher-LLM interactions to support successive optimization rounds^[14], as shown in Figure 3. However, an autonomous experimental feedback loop was not established.

The focus then shifted from coordinating individual research projects to producing reusable synthesis data from the literature. In a further data-oriented extension, Zheng et al. used the ChatGPT Chemistry Assistant to process literature reports into structured MOF synthesis data^[33]. The system extracted synthesis information from published studies and organized it into a form that can be reused by later workflows. It therefore contributed to the data infrastructure needed for more evidence-based agent reasoning, while remaining human-mediated^[33].

Taken together, these studies show that LLM-native agents can decompose research tasks, assign roles across modules, organize literature-derived data, incorporate human feedback, and support optimization or robotic operation. Despite the increasing complexity of workflows, researchers remained the essential interface between the LLM, external tools, experimental equipment, and measured outcomes. The next stage adds persistent external knowledge and specialized computational tools to make agent decisions more scientifically grounded.

LLM-native agents become limited when scientific tasks require evidence beyond the model’s parametric knowledge or the information supplied in the prompt. Stage 2 addresses this limitation by enabling agents to retrieve external knowledge, invoke specialized digital tools, and use the returned outputs to guide subsequent reasoning and decisions. These resources may include literature databases, knowledge graphs, machine-learning models, genetic algorithms, molecular simulations, process models, and deterministic validation tools.

For MOFs and COFs, Stage 2 comprises two closely connected functions. The first is to build machine-actionable knowledge infrastructure that is traceable, updateable, and suitable for reuse. The second is to apply external data and computational tools to material retrieval, property prediction, inverse design, and multi-objective decision-making. The former provides a reliable evidence base, whereas the latter uses that evidence for computational discovery.

Stage 3 begins when automated experimental execution, characterization, and feedback are integrated into the agent’s decision loop. Agent-generated conditions are executed through robotic or laboratory automation systems, and measured outcomes are returned in a machine-readable form to guide subsequent experiments. Human oversight may remain necessary for goal setting, safety, exception handling, and scientific validation, but researchers should not be required to manually interpret and transfer every result between experimental rounds. Full laboratory autonomy and continual retraining of the underlying language model are therefore not prerequisites.

Examples from adjacent materials for MOFs and COFs domains illustrate this architecture. Shi et al. developed the knowledge-driven multi-agent and robotic system MARS as an integrated platform for autonomous materials research^[41]. MARS combines 19 specialized LLM agents, 16 domain-specific tools, hybrid retrieval-augmented generation, and a robotic platform within a hierarchical architecture coordinating literature retrieval, materials design, protocol translation, experimental execution, and data analysis. In optimizing perovskite nanocrystals, MARS used experimental emission data, historical results, and a target wavelength to iteratively update synthesis conditions, tuning the emission wavelength from 523 to 460 nm within ten experimental iterations. The system was also applied to the design and synthesis of water-stable perovskite nanocomposites^[41]. Song et al. developed ChemAgents, a hierarchical multi-agent robotic platform that combines literature analysis, experimental design, predictive modeling, Bayesian optimization, and robotic operation. For metal-organic high-entropy oxygen-evolution catalysts, data from 100 randomly selected and robotically tested compositions were used to update a predictive model and guide Bayesian optimization within a space of 553,401 candidates, yielding a catalyst with an overpotential of 266.1 mV at 10 mA·cm^-2. The workflow established a task-bounded experimental feedback loop and was also transferred to a second robotic laboratory for photocatalytic reactions^[42]. Together, MARS and ChemAgents illustrate experiment-integrated agency in adjacent materials domains, although their reported applications concern perovskites, composites, catalysts, and chemical reactions rather than MOF or COF discovery.

In MOF/COF research, the LLM for Accelerated Synthesis Technique (LFAST) provides a representative example of literature-guided planning, robotic synthesis, automated characterization, and iterative feedback for COF crystallization. A GPT-4o deep-research agent was used to mine, correlate, and validate literature-derived synthesis parameters, thereby defining a screening space under laboratory constraints. Candidate conditions were then executed on a robotic platform, and high-throughput PXRD returned a crystallinity index (CI) that informed subsequent experimental batches^[20], as shown in Figure 10. Using TpPa-SO3H as a benchmark, the study reported an approximately 350% increase in CI relative to previous reports and also enabled the synthesis of the previously unreported COF-2000^[20], as shown in Figure 11. The accompanying PXRD Information File (.pxrdif) further links diffraction data to synthesis metadata, making the results reusable across optimization rounds.

In this workflow, the LLM supported literature-based parameter selection, the robotic platform performed synthesis, PXRD provided the experimental feedback, and CI values were used to inform the selection or refinement of conditions for subsequent screening rounds. However, the precise optimization rule connecting CI to the next round was not fully detailed in the reported study^[20]. More broadly, experiment-integrated Stage 3 agents remain at an early stage in MOF and COF research, with current demonstrations largely focused on well-defined optimization objectives.

LFAST nevertheless represents a supervised and task-bounded form of experimental autonomy, providing an important proof of concept for Stage 3 workflows in this field. Humans still define the objective and constraints, maintain the robotic platform, handle exceptions, and validate chemical conclusions. Its main advance is the automated conversion of experimental measurements into subsequent experimental actions. In the reported workflow^[20], CI served as the primary feedback metric for crystalline order, while complementary evaluation of yield, phase purity, accessible porosity, adsorption performance, and chemical or thermal stability could support more comprehensive, multi-objective optimization in future implementations. Broader autonomy will require such multi-property feedback, automated failure diagnosis, persistent memory, and reliable triggering of subsequent tasks.

Overall, recent studies show that AI agents are beginning to play a more meaningful role in current MOF and COF research. They are no longer limited to simple text assistance, but are increasingly being used to integrate literature reading, database access, simulation, optimization, and in some cases experimental feedback within a more coordinated workflow. Even so, the field is still developing, and most examples so far remain limited in number and scope. A natural next step is to assess how well these emerging strategies extend to a wider range of crystalline porous materials.

In our view, further progress will depend less on using ever larger language models and more on building better scientific support around them. This includes more complete and reusable experimental records, especially data that clearly link composition, topology, synthesis history, characterization, performance, uncertainty, and failed results. It also means that agent outputs need to be checked using chemistry-aware validation, such as charge balance, connectivity, stoichiometry, thermodynamic plausibility, and practical experimental limits, rather than being accepted on the basis of fluent reasoning alone. Another important direction is to make better use of multimodal evidence, since much of the key information in this field is contained in diffraction patterns, adsorption isotherms, spectra, microscopy images, and other non-text data. For closed-loop discovery, standardized instrument interfaces, sample tracking, quality-control records, and machine-readable experimental formats will also be essential, because they allow experimental feedback to be traced and reused across successive cycles. If these pieces come together, AI agents should become not just more capable, but also more trustworthy and more useful in materials discovery.