Pular para o conteúdo principal

Open-Source AI Agent Frameworks

Avançado

Once you've built a few agents by hand, the same plumbing keeps reappearing: a loop that calls the model, runs the tool it asked for, feeds the result back, and stops when the job is done. Agent frameworks package that plumbing — plus state, memory, and multi-agent coordination — so you write less glue. This page is a durable, provider-neutral map of the main open-source options, what they actually give you, and how to choose. Almost all of them are model-agnostic: they work with Claude, GPT, Gemini, and local open-weight models.

What you'll learn
  • Know what an agent framework gives you over a hand-rolled loop — and what it doesn't
  • Recognize the three archetypes: stateful graphs, role-based crews, and minimal loops
  • Use a repeatable procedure to choose one — trading complexity for control
  • Remember that these are mostly model-agnostic — the framework rarely locks you to one provider

What an agent framework actually gives you​

Strip away the branding and a framework is offering you some subset of four things:

  • Orchestration — the control loop. Sequential steps, branches, retries, loops-until-done, and (increasingly) durable execution that survives a crash and resumes where it stopped.
  • Tools — a standard way to declare a function the model can call, validate its arguments, run it, and return the result. The same describe→call→execute→return loop you already know from tool use.
  • Memory & state — somewhere to keep conversation history, scratchpad facts, and retrieved documents across turns, so the agent isn't amnesiac between steps.
  • Multi-agent — patterns for several specialized agents that hand work to each other: a supervisor delegating to workers, or peers collaborating on a task. (Conceptually the same idea as Claude Code's subagents.)

A framework is worth it when you'd otherwise hand-write and maintain all four. It's not worth it when your task is "call the model, run one or two tools, return an answer" — there, the framework is overhead you'll spend time fighting.

The three archetypes​

Frameworks differ less than their marketing suggests. There are really three shapes, and most projects are a variation on one:

  • Stateful graph / workflow. You model the agent as an explicit graph of nodes and edges (or steps and transitions). Maximum control, inspectable state, good for long-running and human-in-the-loop flows. More to learn up front. → LangGraph, LlamaIndex Workflows.
  • Role-based crew. You describe agents by role ("researcher", "writer", "reviewer") and let them collaborate or run a process. Fast to express a multi-agent team; you trade some fine-grained control for the high-level abstraction. → CrewAI, and the conversational multi-agent style of AutoGen.
  • Minimal loop / few abstractions. A thin layer over the model's native tool-calling, with handoffs between a few agents and not much else. Easy to read end-to-end, easy to drop. → OpenAI Agents SDK (the production successor to the experimental Swarm), and the plain loop you write yourself.
Pro tip
  • Start with the simplest thing that works — often a plain tool-calling loop beats a heavy framework.

A quick tour (verify positioning before you commit)​

These are the open projects worth knowing. Named only because each one's real repo was verified; everything volatile lives behind the VerifyNote above.

  • LangGraph — a low-level orchestration framework for stateful, long-running agents modeled as graphs; durable execution and human-in-the-loop are first-class. Usable standalone or with the broader LangChain ecosystem. Model-agnostic.
  • LlamaIndex — started as a data/RAG framework (connectors, indices, retrieval) and now also ships an event-driven Workflows layer for agents. Strong when your agent is fundamentally retrieval over your documents.
  • Microsoft AutoGen — a framework for conversational multi-agent systems. As of mid-2026 it is in maintenance mode; Microsoft directs new projects to a unified successor (Microsoft Agent Framework, merging AutoGen + Semantic Kernel). Check current status before starting.
  • CrewAI — a role-based framework: define agents by role and goal, organize them into Crews (autonomous collaboration) or Flows (event-driven control). Fast path to a multi-agent team.
  • OpenAI Agents SDK — a deliberately lightweight, few-abstractions framework for multi-agent workflows with handoffs. Despite the name it is provider-agnostic (its docs note support for 100+ LLMs), and it is the production-ready successor to the experimental Swarm.
  • The plain agent loop — no framework at all: your own while loop around the model's native tool-calling. The right default for simple agents, and the thing every framework above is ultimately wrapping.

How to choose a framework​

Guided walkthrough1 of 6
  1. One sentence on what the agent does, plus the non-negotiables: latency, cost ceiling, data privacy, whether runs must survive a crash, and whether a human must approve steps.

The thing every framework wraps​

Before reaching for any library, it helps to see the loop they're all built on. This is the entire idea — model decides, you execute, repeat until done:

# Provider-neutral agent loop — the core every framework wraps.
# `model_call` and `run_tool` are yours; swap in Claude, GPT, Gemini, or a local model.

def agent_loop(task, tools, max_steps=10):
    messages = [{"role": "user", "content": task}]

    for _ in range(max_steps):
        # 1. Ask the model what to do next (it sees the tool schemas).
        response = model_call(messages, tools=tools)

        # 2. No tool requested → the model is done. Return its answer.
        if not response.tool_calls:
            return response.text

        # 3. Run each requested tool and feed results back in.
        messages.append(response.as_message())
        for call in response.tool_calls:
            result = run_tool(call.name, call.arguments)
            messages.append({
                "role": "tool",
                "tool_call_id": call.id,
                "content": result,
            })

    return "Stopped: hit max_steps without finishing."

If you can read this, you understand what every framework on this page is doing under the hood. Graphs add explicit state and branching; crews add roles and delegation; minimal SDKs add tidy handoffs — but the heartbeat is always this loop.

Minimal agent system prompt (pairs with the loop above)

You are a task-completing agent with access to tools.

Loop:
1. Think briefly about the next single step toward the goal.
2. If a tool would help, call exactly ONE tool with valid arguments.
3. When you have enough to answer, stop calling tools and give the final answer.

Rules:
- Prefer the fewest tool calls that get the job done.
- If a tool fails, read the error and adjust — do not repeat the same call.
- Never invent tool results; use only what tools actually returned.
- If the goal is impossible with the available tools, say so and stop.

Goal: {one-sentence task}

A note on hype​

No framework here is "the best" — that question is malformed. The graph frameworks win on control and durability; the crew frameworks win on expressing a team quickly; the minimal ones win on readability and low lock-in; and a plain loop wins more often than framework READMEs admit. The right choice is the smallest tool that makes your hardest step clear. As with picking a model, let your own task — not a star count — decide. The same discipline you'd apply with evals applies here: prototype, measure on real cases, keep an exit ramp.

Check yourself

0/3
  1. Your agent makes one model call, runs a single search tool, and returns an answer. What's the right starting point?
  2. You need an agent that runs for a long time, must survive a crash and resume, and lets a human approve certain steps. Which archetype fits best?
  3. A teammate says 'the OpenAI Agents SDK means we're locked into GPT.' Is that right?
Agent frameworks — core ideas
Pressione Enter ou Espaço para virar o cartão. Use as setas esquerda e direita para navegar entre os cartões.Termo exibido.
1 / 6
Key takeaways
  • An agent framework packages orchestration, tools, memory, and multi-agent coordination — adopt one only when you'd otherwise build all four yourself.
  • Three archetypes cover the field: stateful graphs (control/durability), role-based crews (fast teams), and minimal loops (readability, low lock-in).
  • Almost all of these are model-agnostic — they run on Claude, GPT, Gemini, and local models; verify per project rather than assuming.
  • No framework is universally 'best'; pick the smallest tool that makes YOUR hardest step clear, and keep an exit ramp.
  • The plain tool-calling loop is the honest default — and it's exactly what every framework wraps.
  • Maintenance status changes (e.g. AutoGen → a successor); confirm a project is actively maintained at its own repo before committing.

Sources & further reading​

Next​