Disclaimer: I express somewhat of a skeptical view about AI agentic frameworks, but I do see real potential in a number of use cases, I’m excited about the future of the space, and they’re definitely accelerating development of AI applications today. See my last paragraph for a rosier look of the space.
Lately, I’ve been developing several AI apps, and as they’ve grown more complex—especially now that I have multiple AI agents working together—I keep getting pulled toward the promises of agentic frameworks like Langchain, Flowise, CrewAI, Relevance.ai, Vectorshift, Autogen, and others. But after setting up chains, pipelines, and flows in these frameworks, I often find myself reverting to low-code tools like Buildship and Cursor.sh, which allow me to quickly build the frameworks I need.
When I think about the core design patterns where agentic frameworks seem to have the most potential (referencing Andrew Ng’s One Agent for Many Worlds post—where reflection, tool use, planning, and multi-agent collaboration are the common themes), I’ve found that these are fairly straightforward to build. Plus, the underlying functionality these frameworks offer—such as memory, third-party integrations, training, learning, and monitoring—are either:
Not that difficult to build from scratch, and/or
Manageable with features from 3rd party services (e.g., AWS SageMaker, Google’s Vertex AI, etc).
Future of Agentic SystemsOne direction this space could take is more exciting, though. As I discussed in my previous article on Outcome-Driven UX, I can imagine a future where AI understands the outcomes I’m trying to achieve and automatically pulls together the frameworks and multiple AIs needed to accomplish the task. This could further abstract the orchestration of multi-agent systems and make it accessible for non-technical users, transforming how we approach AI-driven app development entirely.
Here are the downsides I’ve encountered:
There’s a high level of abstraction. When I get unexpected results or need to troubleshoot, it becomes much harder to pinpoint the issue. Given my current situation, I’ve found that setting everything up myself is more beneficial for learning—and in most cases, doesn’t take much more time.
Some of the more polished frameworks are not open source, which makes me hesitant to invest heavily in a product that might not be around long-term or that could drastically alter pricing. I’m also unsure whether the services they offer won’t eventually be provided by the big LLM players - see last conclusion at the end.
The space is evolving so rapidly that locking into a specific framework with its own methods or patterns could limit flexibility in the future.
That said, I have to acknowledge that these tools are pretty impressive, especially when it comes to rapid prototyping. The initial versions of my app were fully created using agentic frameworks, allowing me to gather feedback from customers within the first week of development. The no-code interfaces some of these tools offer are fantastic for non-coders, and even for those with a coding background, they provide a solid introduction to multi-agent design patterns, tools, and setup, which can be invaluable for AI app development.
There’s so much out there, and I’ve only experimented with a very small handful. I recently started looking at Langraph, which focuses on creating non-linear, multi-agent workflows. That’s one area where other frameworks have felt confusing or clunky to set up, and Langraph seems to have less abstraction (but with that comes fewer features - e.g. it’s not a no-code/low-code builder like some of the other’s I’ve mentioned), making it more manageable, so it looks promising. OpenAI recently announced Swarm, which I'll also be exploring soon.
I’m sure there will come a time when I adopt one or more of these tools—I just haven’t found the perfect fit yet.
Looking to the FutureWhen considering the future of this space, who knows where it will go. One promising direction that I’d love to see—and as I referenced in my article on Outcome-Driven UX—is the AI understands what I’m trying to accomplish and automatically assembles and orchestrates the necessary AI agents to achieve my goals. Adding this layer of abstraction would dramatically open up the world of developing AI applications to a much larger group of people.
Would love to hear about others' experiences with agentic frameworks, leave in the comments or feel free to send me a message.
Comments