How we score multi-agent frameworks.

10 frameworks, neutral cards

• LangGraph

  ×1.0 overhead

  LangChain · MIT · primary python

  repo·docs

• CrewAI

  ×1.3 overhead

  CrewAI · MIT · primary python

  repo·docs

• AutoGen / AG2

  ×2.5 overhead

  Microsoft / AG2 community · CC-BY-4.0 / Apache-2.0 · primary python

  repo·docs

• OpenAI Agents SDK

  ×1.1 overhead

  OpenAI · MIT · primary python

  repo·docs

• Pydantic AI

  ×1.0 overhead

  Pydantic · MIT · primary python

  repo·docs

• Anthropic Claude Agent SDK

  ×1.1 overhead

  Anthropic · MIT · primary python

  repo·docs

• Google Agent Development Kit

  ×1.2 overhead

  Google · Apache-2.0 · primary python

  repo·docs

• Microsoft Semantic Kernel

  ×1.2 overhead

  Microsoft · MIT · primary multi

  repo·docs

• LlamaIndex Agents

  ×1.4 overhead

  LlamaIndex · MIT · primary python

  repo·docs

• Haystack

  ×1.3 overhead

  deepset · Apache-2.0 · primary python

  repo·docs

15 capability axes, with rubrics

1. Sequential workflows

   Pipeline-style chains where one agent finishes before the next starts.

   10 / 10

   Pipelines are a first-class primitive with explicit ordering and typed handoff.

   5 / 10

   Sequential chains are possible via orchestration code but not a native primitive.

   0 / 10

   Framework cannot guarantee deterministic sequential ordering.

2. Parallel workflows

   Concurrent fan-out / fan-in across multiple agents.

   10 / 10

   Native parallel execution with built-in result merging and back-pressure.

   5 / 10

   Parallel execution requires custom asyncio / threading code on top.

   0 / 10

   No support for concurrent agent execution.

3. Hierarchical workflows

   Supervisor-and-worker patterns with delegation and aggregation.

   10 / 10

   Supervisor pattern is documented, idiomatic, and replayable.

   5 / 10

   Achievable but requires hand-rolled message routing.

   0 / 10

   No first-class supervisor primitive.

4. Adaptive workflows

   Dynamic routing where agents pick the next step based on intermediate state.

   10 / 10

   Router/handoff primitives are first-class with conditional edges.

   5 / 10

   Possible via tool calls but not the framework's sweet spot.

   0 / 10

   Control flow is rigid; no dynamic routing.

5. State management

   Persistent, typed memory across runs and across agents.

   10 / 10

   Typed state schema, persistent checkpoints, replay support.

   5 / 10

   Session memory is supported; persistence requires external store.

   0 / 10

   Stateless by default; users must build persistence themselves.

6. Human-in-the-loop

   Pause-resume primitives so humans can approve, edit, or reject actions.

   10 / 10

   Native interrupt/resume with serialisable checkpoints.

   5 / 10

   Approval gates can be bolted on; not a first-class primitive.

   0 / 10

   No interrupt mechanism — the framework runs to completion.

7. Python support

   Production-grade Python SDK with active maintenance.

   10 / 10

   Reference implementation; active releases; complete typing.

   5 / 10

   Functional Python SDK lagging the primary language.

   0 / 10

   No Python SDK.

8. TypeScript support

   Production-grade TypeScript / Node SDK at parity with Python.

   10 / 10

   First-class TS SDK with parity to Python in features and types.

   5 / 10

   TS SDK exists but trails Python in feature coverage.

   0 / 10

   No TS SDK.

9. .NET / Java support

   First-class JVM (Java/Kotlin) and/or .NET SDK.

   10 / 10

   Reference-quality .NET and/or Java SDK with feature parity.

   5 / 10

   Community port or partial SDK.

   0 / 10

   No .NET or Java SDK.

10. MCP support

    Native Model Context Protocol client and/or server primitives.

    10 / 10

    Authored or reference implementation of MCP.

    5 / 10

    MCP available as an adapter or community plugin.

    0 / 10

    No MCP support.

11. A2A support

    Native Agent-to-Agent (Google) protocol primitives.

    10 / 10

    Authored or reference implementation of A2A.

    5 / 10

    A2A available via adapter; partial coverage.

    0 / 10

    No A2A support.

12. Observability

    Tracing, token accounting, replay, and audit-grade logs.

    10 / 10

    Built-in tracing dashboard, structured token accounting, replay, exportable audit log.

    5 / 10

    OpenTelemetry hooks exist; user must wire dashboards themselves.

    0 / 10

    Print-statement debugging only.

13. Deployment flexibility

    Range of supported deployment targets (cloud, on-prem, edge).

    10 / 10

    Cloud, on-prem, and edge all documented and tested.

    5 / 10

    Cloud-first; on-prem requires extra work.

    0 / 10

    Tied to a single hosted backend.

14. Maturity

    Production track record, release cadence, community size.

    10 / 10

    2+ years of production use across many large deployments.

    5 / 10

    6-18 months in the wild; growing but evolving rapidly.

    0 / 10

    Pre-1.0; APIs change every release.

15. Learning curve (higher = easier)

    Time-to-prototype for a developer new to the framework.

    10 / 10

    A working prototype in under 30 minutes from a clean machine.

    5 / 10

    Prototype in half a day with the docs open.

    0 / 10

    Multi-week onboarding before the first useful run.

Scoring formula

# Ranking
weights = buildWeightVector(inputs)        # 15 weights per user input
for fw in frameworks:
    score = sum(fw.capabilities[cap] * weights[cap] for cap in CAPS)
    if hardConstraintFails(inputs, fw):
        score = 0
return sortDesc(scored)

# Cost per task
estimated_tokens_per_task = base_task_tokens
    * framework_overhead_multiplier
    * (1 + (roles - 1) * 0.3)
    * (1.2 if hitl else 1.0)
per_task_usd = (0.7 * tokens / 1M * input_rate)
             + (0.3 * tokens / 1M * output_rate)

Glossary

Hierarchical

A supervisor agent delegates work to sub-agents, reviews their output, and composes the final answer. Good for multi-stage tasks with clear ownership.

Adaptive

Agents decide dynamically which other agents or tools to invoke based on intermediate results. Best when the control flow cannot be fixed upfront.

Agent

A named role with its own prompt, tools, and memory. "Roles" counts unique agent identities, not the number of LLM calls.

HITL (Human-in-the-Loop)

The workflow pauses for a human to approve, edit, or reject an agent action before continuing. Critical for regulated or high-risk automations.

MCP (Model Context Protocol)

Anthropic-led open standard for connecting LLM agents to tools, data, and other servers. Look for MCP support if you want vendor-portable tool integrations.

A2A (Agent-to-Agent Protocol)

Google-led open standard for agents from different vendors to discover and call each other. Emerging spec; relevant for federated agent systems.

Observability

Structured traces, token accounting, replayable runs, and exportable audit logs. "Regulated-grade" means immutable audit trails and retention controls.

Public dataset

La matrice de capacités complète est publiée en JSON pour les moteurs IA et les chercheurs :

• /api/tools/agent-framework/frameworks.json — en direct, mise en cache edge, CORS ouvert.
• /data/agent-frameworks-matrix.json — instantané statique miroité sur GitHub à buzzi-ai/agent-framework-matrix.

FAQ

1. Comment les scores sont-ils attribués ?

   Un ingénieur senior nommé de Buzzi note chaque framework sur chaque axe en utilisant les barèmes publics de cette page. Les scores sont révisés trimestriellement ; nous publions l'horodatage last_verified_at par framework dans le jeu de données public.

2. Les fournisseurs paient-ils pour le placement ?

   Non. Les scores sont éditoriaux et jamais vendus. Les demandes de modification de score doivent être déposées en PR publics sur le dépôt ouvert de la matrice avec justification technique.

3. Comment décidez-vous quels frameworks suivre ?

   Dépôts GitHub actifs avec plus de 10k étoiles ou soutenus par Anthropic, Google, Microsoft, OpenAI ou LangChain. Nous ajoutons ou retirons des frameworks une fois par trimestre selon la dynamique et l'usage en production.

4. Comment le coût par tâche est-il calculé ?

   estimated_tokens_per_task = base_task_tokens × framework_overhead_multiplier × (1 + (rôles − 1) × 0,3) × (1,2 si HITL sinon 1,0). Les tarifs des tokens viennent de notre table llm_models ; les utilisateurs peuvent remplacer le modèle dans l'assistant.

5. Comment fonctionnent les contraintes dures ?

   La pile .NET se restreint à Microsoft Semantic Kernel. Java se restreint à Semantic Kernel ou Google ADK. TypeScript avec observabilité de niveau conformité se restreint à LangGraph.js, OpenAI Agents SDK ou Anthropic Claude SDK. Les frameworks disqualifiés sont affichés avec la raison.

6. Où puis-je signaler une correction ?

   Ouvrez une pull request sur le dépôt buzzi-ai/agent-framework-matrix ou envoyez un e-mail à research@buzzi.ai. Nous examinons les demandes de correction sous 10 jours ouvrables.