Compare
The same six axes across every evaluated framework. Read the columns, not just the total — two frameworks can land three points apart and be built for opposite jobs.
| Axis | Weight | Pydantic AI | LangGraph | OpenAI Agents SDK | CrewAI |
|---|---|---|---|---|---|
| Control flow & multi-agent compositionSequences, routers, graphs, parallel work, loops, handoffs, and whether control is deterministic or model-routed. | 20% | 3/4 | 4/4 | 3/4 | 3/4 |
| State, memory & durabilityTyped state, session vs long-term memory, checkpoints, persistence, resume/restart, retry, and idempotency semantics. | 20% | 4/4 | 4/4 | 3/4 | 4/4 |
| Tools, integrations & interoperabilityTyped tool contracts, structured output, streaming, provider/plugin adapters, and native or official protocol support. | 15% | 4/4 | 3/4 | 4/4 | 3/4 |
| Reliability, safety & human controlsTimeouts, retries, cancellation, guardrail enforcement, permissions, isolation, approvals, interrupts, and secret boundaries. | 15% | 4/4 | 4/4 | 4/4 | 3/4 |
| Observability & evaluationRun/trace identity, model/tool/handoff events, logs, metrics, cost visibility, replay, evals, and test fixtures. | 15% | 4/4 | 3/4 | 4/4 | 3/4 |
| Developer experience, portability & maintainabilityClean install, canonical quickstart, public API/types, supported runtimes, deployment path, versioning, and migrations. | 15% | 4/4 | 3/4 | 3/4 | 3/4 |
| Architecture score | 100% | 95 | 89 | 86 | 80 |
Highlighted cells mark the leading score on that axis. A score of 4 requires evidence of failure, recovery or enforcement behaviour — not just that a feature exists.