Module frameworks 

Evaluating AI frameworks for agentic AI use.

Where languages profiles the language an agent writes in, this module profiles the AI framework an agent builds with — the library it must discover, drive, and debug autonomously. Same four axes, framework-flavored:

• token efficiency — how many tokens a working model/pipeline costs (API verbosity, config boilerplate, import surface).
• determinism — seeded-run reproducibility, version stability, and whether artifacts (checkpoints, graphs) are byte-stable.
• reliability — when the agent misuses the API, does it get an early structured error (shape checks at graph build) or a runtime tensor explosion three layers deep?
• safety — does loading/running third-party artifacts execute arbitrary code (pickle!), and is the compute surface effect-gated?

Plus one framework-specific axis the others don’t need:

• discoverability — can an agent learn the surface from the framework itself (machine-readable schemas/ontology, introspectable ops, stable programmatic docs) instead of scraping prose?

Profiles are curated 0.0–1.0 static judgments with evidence, like the language profiles — deterministic, serializable, comparable.

use agentic_eval::frameworks::{profile, rank_frameworks, Framework};
let torch = profile(Framework::PyTorch);
assert!(torch.evidence.len() >= 3);
let ranked = rank_frameworks();
assert!(ranked[0].fitness() >= ranked[ranked.len() - 1].fitness());

Structs

FrameworkComparison

Compare two frameworks: positive deltas mean a fits agentic use better.

FrameworkProfile

A curated agentic profile of an AI framework: the four shared axes plus framework-specific discoverability, with evidence.

Enums

Framework

AI frameworks with curated agentic profiles.

Functions

compare_frameworks

Compare framework a against baseline b across all five axes.

profile

The curated profile for fw (static, documented judgments — see module docs).

profiles

Profiles for all frameworks, in Framework::all order (deterministic).

rank_frameworks

All profiles ranked best-first by FrameworkProfile::fitness (stable order on ties).