Agent frameworks, evaluated against evidence you can check.

Every score here points at a file, at a pinned commit, containing the thing it claims. The build fails if the citation does not hold. Popularity is reported separately and never touches the score.

frameworks evaluated

verified citations

research records

Rubric v1.0.0 · citations verified 18 Aug 2026 · repository snapshot 18 Aug 2026

Ranked by architecture score

How scoring works →

95/100

Pydantic AI

Pydantic · Python · pydantic/pydantic-ai

Type-first agent framework that treats durability and OpenTelemetry as architecture, not integrations.

control 3
state 4
tools 4
reliability 4
observability 4
developer 4

89/100

LangGraph

LangChain · Python · langchain-ai/langgraph

Low-level graph runtime for long-running, stateful agents, with checkpointing as a first-class primitive.

control 4
state 4
tools 3
reliability 4
observability 3
developer 3

86/100

OpenAI Agents SDK

OpenAI · Python · openai/openai-agents-python

Small primitive set — agents, handoffs, guardrails, sessions — with tracing built in rather than bolted on.

control 3
state 3
tools 4
reliability 4
observability 4
developer 3

80/100

CrewAI

CrewAI Inc. · Python · crewAIInc/crewAI

Role-based crews for delegation-shaped work, plus a decorator flow layer and event-driven checkpointing.

control 3
state 4
tools 3
reliability 3
observability 3
developer 3

Two layers, never blended

An architecture score built only from evidenced axes, and a repository snapshot shown beside it. Stars cannot make a design better. The score function is structurally unable to read them.

Citations pinned to a commit

Each claim names a path and the strings that must appear in it. A verifier fetches that file at the reviewed SHA and fails the build when the text is missing.

Unknown is not zero

An axis without sufficient evidence stays Unknown, and a record with any Unknown axis publishes no aggregate at all rather than a flattering guess.

Research

All 17 records →

Papers, memos and threads kept deliberately separate from the directory. They are not frameworks, carry no score, and are never counted as evaluated tools.