TL;DR: If you just need tool-calling agents, start with the OpenAI Assistants API. For complex reasoning without orchestration overhead, Claude's tool_use is hard to beat. For teams wanting zero-infrastructure scheduled agents, Nebula handles it end-to-end. For visual builders, check out Dify.ai.
Why Skip the Framework?
You wanted to build an AI agent that sends Slack summaries every morning. Somehow you ended up debugging LangChain dependency conflicts at 2am.
Sound familiar?
Frameworks like LangChain, CrewAI, and AutoGen are powerful. They handle complex multi-agent orchestration, custom retrieval pipelines, and research-grade workflows. But here is the thing: most agent use cases don't need any of that.
A scheduled task that reads emails, summarizes them, and posts to Slack? That does not need a 50,000-line framework. A support bot that looks up docs and drafts replies? Same story.
The anti-framework movement is gaining real momentum. Andrej Karpathy's recent autoresearch project proved that a tight loop of a few hundred lines of Python can run autonomous research experiments -- no orchestration layer required.
So what are the alternatives? Here are five tools that let you build production AI agents without adopting a framework.
Quick Comparison
| Feature | OpenAI Assistants | Claude tool_use | Composio | Dify.ai | Nebula |
|---|---|---|---|---|---|
| Setup complexity | Low | Low | Medium | Low | Low |
| Model flexibility | OpenAI only | Anthropic only | Any LLM | Multi-model | Multi-model |
| Built-in integrations | Limited | None | 250+ | Moderate | 1,000+ |
| State / memory | Thread-based | None built-in | None built-in | Conversation memory | Persistent memory |
| Scheduling | None | None | None | Basic | Built-in cron |
| Self-hostable | No | No | Yes | Yes | No |
| Pricing | Usage-based | Usage-based | Free tier + paid | Free self-hosted | Free tier available |
1. OpenAI Assistants API
Best for: Developers already in the OpenAI ecosystem who want tool-calling agents fast.
The Assistants API gives you a managed agent runtime with built-in tools: code interpreter, file search, and function calling. You define your tools as JSON schemas, and the model decides when to call them.
The thread-based memory system handles conversation state automatically. You create a thread, add messages, and run the assistant -- it manages the tool-calling loop for you.
Strength: Fastest path from zero to a working tool-calling agent. Code interpreter lets your agent write and execute Python on the fly.
Weakness: You are locked into OpenAI models. If GPT-4o has a bad day or pricing changes, you cannot swap to Claude or Gemini without a rewrite.
Pricing: Usage-based. You pay for model tokens plus tool usage (code interpreter sessions, file storage).
2. Anthropic Claude tool_use
Best for: Complex reasoning tasks where decision quality matters more than integration breadth.
Claude's tool use takes a different approach. You pass tool definitions in your API call, Claude decides which tools to call and with what parameters, and you execute the tools yourself. It is a lower-level primitive than Assistants API, but that is the point.
You get Claude's best-in-class reasoning for deciding when and how to use tools, while keeping full control over execution. Structured outputs via tool_use mean you get clean JSON back every time.
Strength: The reasoning quality for multi-step tool chains is arguably the best available. Claude thinks carefully about tool sequencing rather than firing everything at once.
Weakness: No orchestration layer. You build the loop, manage state, handle retries, and wire up every integration yourself. This is a building block, not a product.
Pricing: Standard Claude API pricing. You pay per token for input and output.
3. Composio
Best for: Developers who need agents connected to many SaaS tools without building every integration from scratch.
Composio solves one of the most annoying parts of building agents: integrations and authentication. It provides 250+ pre-built tool connections (GitHub, Slack, Gmail, Jira, Notion, and more) with managed OAuth and API key handling.
You bring your own LLM -- OpenAI, Claude, Gemini, open-source models -- and Composio handles the tool execution layer. It plugs into frameworks too, but works perfectly fine with plain API calls.
Strength: Auth management alone can save you days. Instead of building OAuth flows for every service, Composio handles token refresh, scoping, and error handling.
Weakness: You still need to wire up the LLM, manage conversation state, and build your own scheduling or triggering logic. Composio is the integration layer, not the full stack.
Pricing: Free hobby tier available. Paid plans start around $29/month for higher usage limits.
4. Dify.ai
Best for: Teams who want to design agent workflows visually without writing orchestration code.
Dify is an open-source platform for building LLM-powered applications through a visual drag-and-drop interface. You can design chatbots, RAG pipelines, and agent workflows by connecting nodes -- LLM calls, tool integrations, conditional logic, and knowledge retrieval.
The open-source version is genuinely full-featured. You can self-host on your own infrastructure with Docker, which makes it appealing for teams with compliance requirements.
Strength: The visual builder makes agent logic transparent and debuggable. Non-engineering team members can understand and modify workflows. The self-hosted option is a real differentiator.
Weakness: Custom code is a second-class citizen. When you need to do something the visual builder does not support, workarounds get awkward. The developer experience favors no-code over code-first.
Pricing: Free and open-source for self-hosted. Cloud plans start at $59/month for teams.
5. Nebula
Best for: Developers who want scheduled, multi-step AI agents running in production without managing infrastructure.
Nebula takes a different approach from the tools above. Instead of giving you building blocks and leaving you to assemble them, it provides a managed platform where agents run on schedule, connect to 1,000+ apps (via Pipedream), and maintain persistent memory across runs.
You define what the agent should do -- check emails, summarize PRs, post reports to Slack -- and Nebula handles the execution loop, retries, state management, and integrations. It supports multi-agent delegation, so complex workflows can be broken into specialized sub-agents.
Strength: Zero infrastructure overhead. Agents run on cron schedules, remember context between executions, and connect to your existing tools without building integration code. The email-native capability (agents can send and receive email) is unique.
Weakness: Newer platform with a smaller community compared to OpenAI or Anthropic ecosystems. Less flexibility for highly custom agent architectures.
Pricing: Free tier available for getting started.
When You Should Still Use a Framework
To be fair, frameworks exist for good reasons. Consider LangGraph, CrewAI, or AutoGen when you need:
- Custom retrieval pipelines with fine-grained control over chunking, embedding, and reranking
- Research-grade multi-agent systems where agents negotiate, debate, or collaborate
- Full control over every LLM call, prompt template, and retry strategy
- Complex graph-based workflows with conditional branching that goes beyond linear execution
Frameworks give you maximum control. The trade-off is maximum setup time and maintenance burden.
The Verdict
For most agent use cases -- scheduled tasks, API integrations, email and Slack automation, support workflows -- you do not need 10,000 lines of framework code.
Pick the tool that matches your complexity level:
- Prototype fast with one model? Start with OpenAI Assistants API.
- Need the best reasoning for tool decisions? Use Claude tool_use and build your own loop.
- Connecting to many SaaS tools? Composio saves you the integration headache.
- Want a visual builder your whole team can use? Dify.ai is the move.
- Need agents running on schedule with zero DevOps? Nebula handles the full lifecycle.
The right tool is the one where you spend time on what the agent does, not on the infrastructure it runs on.
Top comments (0)