Developer Perspective of AI Agent Frameworks
2025 is the year of agents. If you are a AI developer these days, you are heavily involved in developing agents. An year back from a developer standpoint, we had 2 frameworks available to use or you can build your own framework. Now you have at least 25 different frameworks. None of these are production ready, out of the box.
At Transilience AI we have been building agentic applications for a quite a long time and hope our experiences can help the community as the feud between agentic frameorks escalate.
As a developer I had hoped for same sanity in the agent frameworks , some standards so I an focus on AI , keep the joy of AI development and focus less on frustrations related to the frameworks them selves. It was frustrating to see each of the frameworks take funding and nudge us to use their own paid versions of the frameworks. What is more frustrating to see is the public feuding of the frameworks against each other.
What is important from a developer stand point ?
- Cognitive overload: We are AI engineers, not a framework engineer. The more the framework itself occipies our mind when maintaining out code, we are losing out on maintaining our edge related to makign our application awesome and making our user experience better.
- Debuggability: Looking at logs is important. Tracing is important, evals are important. If we have to change our application code and the way our applications have to behave to enable tracing and worst yet get fored to use the frameworks paid SaaS, we are on the losing end
- Extensibility: There are several agentic patterns that our application want to use and extending the framework to use the patterns and make our own custom patterns is paramount. The ease at which we can test these patterns will help us to iterate faster on the agentic capabilities of the application.
So What is the Answer ?
An year back, we developers used to have the same frustration with models. As new models come out and as they up their game, switching models and prompts was frustrating and trying to remember which prompt engineering and prompt formatting techniques are better with what models made us insane. Models getting better and less reliant on specific prompt language helped. What has helped more is building our own little wrappers so we can switch models and plug and play models and parameters. There is a new model X that came out and has better capabilities ? One line change.
One solution that we are executing at Transilience AI is a notion of our agentic applications using "workflows" that we can plug in multiple agentic frameworks into. OpenAI handoffs is good for a security use case, one line change.
Of course we cannot make our wrapper consider all frameworks out there, if we did that then we would ourselves become the wrapper hell to maintain. So which frameworks to maintain ?
- Frameworks from foundational models : We cannot escape from Agentic frameworks such as OpenAI, Google or Anthropic (MCP or their own agent framework when they come around to it)
- Build your own (Pydantic) : Pretty much every framework out there relies on pydantic for passing function definitions so you have to use pydantic to build on your own. Along with that Pydantic has their own framework
You might ask, why dont we just use one framework ? Well a) vendor lock in and more importantly b) a third party abstraction framework brings in tons of other third party tools and maintaining abstractions from 3-4 agent framework providers is the most we can do from a cognitive overload perspective.
Twitter conversation
And there is no getting away from agent frameworks from foundational m,odel folks as its their "SDK". They would release their core functionality through their agents, say Google ecosystem support through Google ADK etc.
So in summary maintain small set of frameworks and make it simple to stich workflows from the best of breed
- OpenAI agents are very good at hand offs
- Google agents are good at declarative multi agents
- Pydantic agents are good at building DAG on top of any agents
The above frameworks are good enough for most production use cases.
Transilience AI backend team ( Alessio Mauro , SHASHWAT SHARMA , Smritika Sadhukhan ✨ )