I am very curious about this as well. I'm looking for something that does really well with workflows that require 20 plus steps including a couple while loops and user verifications, but also something simple like a chat bot with access to MCP servers and tools. I need to be able to use Gemini directly, openAI directly, and openrouter at various steps.
Right now I am having trouble deciding whether it's better to just write my own harness, or use langchain, or something else.
You can do all of that with current frameworks depending on how deep you go. The issue people are finding is that one size does not fit all and your specific workflows may be better suited for a lower level.
What I find impossible to judge is whether me choosing the harness that works best for me and the way I like to work will limit the quality of the LLM output.
In this case, given the complexity of LangChain I don't know if it would burn a lot of tokens or confuse the LLMs with context creep due to the large env and tooling compared to something much simpler, or would something much simpler burn a lot of tokens and stutter in execution compared to LangChain, because they have a lot of middleware optimizations that I would have to relearn the hard way? Or are both those strategies off the mark and there is a better tool for the job I am not even thinking about?
It gets a bit expensive in terms of both time/effort and money to experiment on a full workflow rather than specific steps which is quite easy in openrouter, hence my curiosity as to others' experiences.
I like it quite a bit. Imo it's sort of like the "Flask" of frameworks. It's pretty easy to get started with and has a pretty pluggable ecosystem where you can choose models, providers, tools without much lock-in.
It has AWS weight behind it (for better or worse) and has a slight skew towards AWS solutions but it's still trivial to skip those.
Krew.ai is another one not on your list. It's fairly comprehensive but it was too "enterprisy" and too heavy on proprietary jargon for all the components imo. I think the tech is ok but the docs and terminology is very dense.
I also briefly looked at LangGraph and it's fairly straight forward and provides a lot of the same abstractions and interfaces as Strands but it felt too easy to accidentally miss them and end up reinventing the wheel. Put another way, it was easy to accidentally miss the abstraction and wade into DIY primitive territory. That may be a feature to some, but I thought it would be a distraction for quick adoption at a small company.
When I checked a few months ago, LangGraph, Krew, and OpenAI Agents SDK had better out of the box integration with common observability/monitoring/tracing solutions. Strands has pretty robust Otel support but not all providers had implemented the Otel AI spec at the time so it required some adapters/wrapper code (plenty of community examples online, just slightly less plug and play)
At the time, I don't think Claude Agents SDK existed yet or it was very new. I skipped OpenAI SDK because we wanted a multi-vendor solution. Pydantic Agents was too new so I skipped that
its good that there is a competition in framework space, but does anyone have holistic view or opinions about their differences and where they shine?
For example,
* there is LangChain and LangGraph - used a lot, but framework bloat is hated as well
* mastra - for typescript projects
* pydantic, agno, strands, openai agents sdk, claude agents sdk, and so on and on and on
I am very curious about this as well. I'm looking for something that does really well with workflows that require 20 plus steps including a couple while loops and user verifications, but also something simple like a chat bot with access to MCP servers and tools. I need to be able to use Gemini directly, openAI directly, and openrouter at various steps.
Right now I am having trouble deciding whether it's better to just write my own harness, or use langchain, or something else.
You can do all of that with current frameworks depending on how deep you go. The issue people are finding is that one size does not fit all and your specific workflows may be better suited for a lower level.
What I find impossible to judge is whether me choosing the harness that works best for me and the way I like to work will limit the quality of the LLM output.
In this case, given the complexity of LangChain I don't know if it would burn a lot of tokens or confuse the LLMs with context creep due to the large env and tooling compared to something much simpler, or would something much simpler burn a lot of tokens and stutter in execution compared to LangChain, because they have a lot of middleware optimizations that I would have to relearn the hard way? Or are both those strategies off the mark and there is a better tool for the job I am not even thinking about?
It gets a bit expensive in terms of both time/effort and money to experiment on a full workflow rather than specific steps which is quite easy in openrouter, hence my curiosity as to others' experiences.
> Strands
I like it quite a bit. Imo it's sort of like the "Flask" of frameworks. It's pretty easy to get started with and has a pretty pluggable ecosystem where you can choose models, providers, tools without much lock-in.
It has AWS weight behind it (for better or worse) and has a slight skew towards AWS solutions but it's still trivial to skip those.
Krew.ai is another one not on your list. It's fairly comprehensive but it was too "enterprisy" and too heavy on proprietary jargon for all the components imo. I think the tech is ok but the docs and terminology is very dense.
I also briefly looked at LangGraph and it's fairly straight forward and provides a lot of the same abstractions and interfaces as Strands but it felt too easy to accidentally miss them and end up reinventing the wheel. Put another way, it was easy to accidentally miss the abstraction and wade into DIY primitive territory. That may be a feature to some, but I thought it would be a distraction for quick adoption at a small company.
When I checked a few months ago, LangGraph, Krew, and OpenAI Agents SDK had better out of the box integration with common observability/monitoring/tracing solutions. Strands has pretty robust Otel support but not all providers had implemented the Otel AI spec at the time so it required some adapters/wrapper code (plenty of community examples online, just slightly less plug and play)
At the time, I don't think Claude Agents SDK existed yet or it was very new. I skipped OpenAI SDK because we wanted a multi-vendor solution. Pydantic Agents was too new so I skipped that
I've experience using it with clients on several small to large projects. It's has advantages and disadvantages, as every framework.
Clients choose it because it's EU-based company.
Would you have used something else without that constraint?
RubyLLM is where its at.
Why Ruby? It seems to be a Python framework.
Could have easily been EmeraldLLM… (iykyk)
deepset? more like deadseek.
no, thanks, never going to use that.
> deepset? more like deadseek. Just curious, why say that ?
my bet is propaganda