A single language model call can summarise a document, draft an email, or classify support tickets. But most real business workflows chain multiple steps: read a contract, extract key obligations, cross-reference against a database, generate a structured briefing. For these multi-step processes, a single LLM prompt is rarely enough — and that is exactly the problem multi-agent orchestration frameworks solve.
Three open-source frameworks have matured to the point where they are genuinely viable for SMB production use: LangGraph, CrewAI, and AutoGen. All three run entirely on-premise, all three connect to Ollama with local open-weight models, and none of them require a cloud API subscription. As @PythonDv wrote on X this week: "You don't need to spend a single dollar to build a production AI system in 2026" — listing LangGraph and CrewAI as the orchestration layer (source). The post resonated widely because it captures where the open-source ecosystem stands right now.
The practical question for businesses is not whether these frameworks work — it is which one fits a particular use case, team, and infrastructure.
LangGraph: Precise Control for Production-Grade Workflows
LangGraph, built by the LangChain team, models agent workflows as a directed graph. Each node is a processing step or agent call; edges define transitions, branching conditions, and loops. This structure gives developers fine-grained control over exactly what happens at each step — and what happens when something goes wrong.
Strengths:
- Full control over multi-step, branching processes — no hidden magic
- Explicit state management: every step can be logged, inspected, and replayed for debugging
- Battle-tested in production; large community, well-documented
- Strong fit for processes where auditability matters (compliance, finance, legal)
Limitations:
- Steeper learning curve than CrewAI; more configuration code for simple tasks
- Tightly coupled to the broader LangChain ecosystem
Ollama integration: LangGraph connects to Ollama via the ChatOllama class. Setting base_url: http://localhost:11434 is all that is needed to route all LLM calls to a local model. As reported by practitioners, Llama 3.3 70B on a Mac Studio M3 Ultra (192 GB unified memory) delivers approximately 18–28 tokens/s — fast enough for interactive business workflows.
CrewAI: Role-Based Agents for Rapid Prototyping
CrewAI organises agents around an intuitive metaphor: a "crew" of agents each with a defined role, goal, and set of tools. A research agent, a writing agent, a quality-review agent — each has a job, and CrewAI handles coordination. The role abstraction makes workflows easy to reason about even for people who are not deeply familiar with AI infrastructure.
Strengths:
- Simplest API of the three frameworks — working prototypes in hours, not days
- Role metaphor maps naturally to existing team structures and business processes
- Parallel agent execution is straightforward to configure
- Low overhead for straightforward task decomposition
Limitations:
- Less fine-grained control over execution flow compared to LangGraph
- Complex conditional logic can become harder to manage at scale
Ollama integration: CrewAI supports Ollama natively through the ollama/ model prefix. No additional LangChain configuration layer required, which keeps the setup clean and minimal.
AutoGen (Microsoft Research): Conversational Collaboration Between Agents
AutoGen models multi-agent workflows as a conversation between specialised agents that exchange messages until they reach a result — similar to a structured expert discussion. Since version 0.4, AutoGen supports OpenAI-compatible endpoints, which means Ollama works as a drop-in local backend.
Strengths:
- Natural fit for tasks requiring iterative back-and-forth: code review, research synthesis, multi-perspective analysis
- Native human-in-the-loop support — a human can intervene in the agent conversation at any point
- Clean VS Code integration for development teams already in a Microsoft toolchain
Limitations:
- Conversational approach can consume significantly more tokens per task than graph-based frameworks on longer workflows
- Debugging complex agent conversations takes more effort than inspecting a deterministic graph
Ollama integration: AutoGen accepts any OpenAI-compatible endpoint. Ollama is configured as a local proxy in the model_client settings — no code changes needed beyond the endpoint URL.
Which Framework for Which Business Use Case?
| Use case | Recommendation | Why |
|---|---|---|
| Multi-step document processing | LangGraph | Precise flow control, reliable logging |
| Content production, research pipelines | CrewAI | Role metaphor, fastest to build |
| Iterative analysis, code review | AutoGen | Conversational logic, human-in-the-loop |
| First pilot with a non-developer team | CrewAI | Lowest barrier to entry |
| Production-critical, auditable workflows | LangGraph | Reproducibility, full state inspection |
For most SMBs starting out, CrewAI is the right choice for a first pilot — it produces visible results quickly, and the role-based structure maps well to how teams think about tasks. Once a workflow moves toward production and reliability matters, LangGraph's additional configuration overhead pays for itself through predictability and debuggability.
Data Sovereignty: No Cloud, No Data Transfer, No Compliance Risk
With a fully local multi-agent stack, nothing leaves the company network — not the documents being processed, not the queries sent to the language model, not the intermediate results passed between agents. This is the architectural difference from cloud-based agent services like OpenAI's Agents API or similar offerings.
For businesses operating under GDPR, this has concrete implications:
- No data processing agreement with a third-party model provider required
- No exposure to cloud-side security incidents at external vendors
- No risk of sensitive business data appearing in third-party model training pipelines
Based on our reading of the EU AI Act, businesses that run exclusively local open-weight models without third-party API calls are in a significantly better regulatory position regarding the deployer obligations under Article 26 that take effect in August 2026. See our data sovereignty overview for what this means in practice.
Recommended Hardware and Models
Based on community-reported benchmarks:
- Entry level (Llama 3.2 11B, Gemma 3 12B, Mistral Small 4 22B): MacBook Pro M3/M4 with 32–36 GB unified memory; 20–45 tokens/s reported
- Mid-range (Llama 3.3 70B, Qwen 2.5 72B): Mac Studio M3/M4 with 64–192 GB; 18–30 tokens/s reported
- Heavy workloads (70B Q8 or large MoE models): Mac Studio M3/M4 Ultra 192 GB or NVIDIA RTX 6000 Ada (48 GB VRAM)
For typical SMB workflows — document analysis, internal knowledge retrieval, report generation — a model in the 14B–32B range is sufficient and deployable on a single Mac Studio.
Getting Started
The path from experimentation to production is straightforward:
- Install Ollama (ollama.ai) and pull your chosen model:
ollama pull llama3.3orollama pull mistral-small - Install the framework:
pip install langgraph/pip install crewai/pip install pyautogen - Define a concrete first use case — internal document search, contract analysis, email triage — rather than an abstract proof of concept
- Measure and iterate: document quality, latency, and resource consumption from the first run
Our local AI for business overview covers common deployment patterns and typical hardware configurations for European SMBs. If you want hands-on support in selecting the right framework and deploying a first production workflow on your own infrastructure, our pilot project programme covers exactly that — from framework selection to live deployment, with your data staying on-premise throughout.