Generative models for polymer structure optimization, modification and property prediction
PolyAgent has 9 tools with moderately good naming (verb-based, domain-specific) and reasonable parameter schemas, but suffers from significant gaps in description quality and lacks critical error-handling guidance. Tool descriptions range from adequate (sa_score_smiles) to vague (decode_latents lacks context on what 'optimized_latents' structure is). Most parameters lack comprehensive descriptions of their formats, ranges, and constraints. Output schemas are partially documented through code comments and return statements, but LLMs would struggle to understand what fields to expect without parsing the implementation. No evidence of per-parameter constraint documentation (ranges, patterns, enums where appropriate). Error handling is minimal, no guidance on what to do if optimization fails, model loading fails, or constraints are unsatisfiable. The codebase uses session-based state management (MODEL_SESSIONS) which is reasonable for the domain, but this stateful design is not exposed in tool descriptions, forcing agents to guess how sessions work. Security is adequate (no credentials in params), but no permission checks or audit trails are visible.
Decode optimized latent vectors into polymer SMILES structures and format results with property values
Load OpenMacromolecularGenome multiproperty model with encoder, decoder, and property predictor
Load OpenMacromolecularGenome single property model for optimization
Optimize latent vectors under constraints to achieve target property values
Complete pipeline: Load model, optimize latent vectors for target property value, and decode to polymers
Complete pipeline for single property models: Load single property model, optimize latent vectors for target property value, and decode to polymers
Missing or inadequate parameter descriptions for complex types. Parameters like 'optimized_latents', 'constraint_spec', and 'constraints' are described as generic objects without explaining their structure, expected fields, or format. LLMs cannot determine what to pass without reading implementation code.
Missing constraint documentation for numeric parameters. Parameters like 'num_seeds', 'num_steps', 'step_size', 'tolerance', and 'lambda_constraint' lack min/max bounds, valid ranges, or guidance on typical values. Default values exist but ranges are not documented.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 53 | 2026-07-28+ | v2 |
Optimize latent vectors under constraints for single property models
Predict polymer properties from SMILES using TransPolymer models
Calculate synthetic accessibility (SA) scores for a list of SMILES strings
Inadequate tool descriptions for disambiguation. 'optimize_single_property_latents' and 'optimize_single_polymer_property' both optimize polymer properties but differ in scope, one returns raw latents, the other returns complete results. Descriptions do not clearly distinguish when to use each.
Missing output schema documentation. Tools like 'optimize_latents' and 'optimize_single_property_latents' do not document what fields agents should expect in the response. The code returns a dict but the structure is inferred from implementation rather than declared.
No error recovery guidance. Tool descriptions do not explain what happens if optimization converges slowly, constraints are unsatisfiable, or models fail to load. No recovery patterns provided.
Session-based state (MODEL_SESSIONS) is not documented in tool descriptions. Agents must infer from code that load_model() must be called before optimization tools, but descriptions do not state this dependency.
Missing enum constraints. The 'property' parameter in 'predict_properties' correctly uses enum ['Eea', 'Egb', 'EPS', 'PE_I', 'OPV'], but other tools that accept property names (e.g., 'target_property' in optimization tools) are free-form strings with no enum constraint.
Inconsistent parameter naming across tools. Some tools use 'target_property' and 'target_value', others use 'constraint_spec' with nested structure. No consistent interface for specifying optimization targets.