In detail
What Jerry Liu argues
Liu describes LlamaIndex’s move toward deeper work on document understanding and production context. In the interview at 38:28, he explains that abstractions were stabilizing while the need for high-quality context continued to evolve. This is his account of where his company could add value, rather than evidence that frameworks have become unnecessary for every team. Earlier in the interview at 1:56, he describes when a framework layer does earn its place: while a space is still being figured out and applications need a lot of flexibility, a framework supplies the core abstractions that make certain things easier to build.
What a direct implementation looks like
The direct alternative can be small. Vercel CTO Malte Ubl described one in episode 39 at 21:47: give the model its tools and a goal, allow “up to 10 turns using a tool,” and let it work out the sequence. A loop like that is a few dozen lines of your own code, and every step is yours to log.
Questions to test in your own project
Our practical recommendation is to compare a framework with a small direct implementation of the same workflow:
- Can you inspect tool calls and state when the agent fails?
- Can you resume interrupted work without repeating an action? This is where frameworks often earn their keep: LangGraph’s persistence layer, for example, saves state through checkpointers and documents fault tolerance. Check whether your own loop would need to rebuild that.
- Does the framework simplify the branching or coordination your application needs?
- Can you change retrieval and evaluate the resulting answers without fighting its abstractions?
Anthropic’s engineering team gives similar advice in Building effective agents: frameworks make it easy to start, but they add layers that can obscure the underlying prompts and responses and make failures harder to debug. Its suggestion is to start with direct model API calls, since many patterns take a few lines of code, and if you adopt a framework, to understand what it does underneath. Use the comparison to decide whether the saved implementation work justifies the dependency and its constraints. A simple workflow may need little orchestration; a complex one may benefit from a framework. Neither choice removes the need to test the data supplied to the model.
For the current options and their trade-offs, see AI agent frameworks compared. For another perspective, compare the agent-stack and context arguments. They address different parts of building a production system and can both be useful.
Disclosure: Episode 39 aired while Chain of Thought was Galileo’s podcast and Conor worked there.
Hear it from the guest
“When there's a space where things are still being figured out, and then there's kind of a lot of flexibility in the end application, and that's where there's kind of value in a framework layer because the framework provides you know the core abstractions to make it easier to build certain things”
“Because the abstractions were solidifying, but the need for high-quality production context was still evolving.”
“You tell the LM, okay, you get to make up to 10 turns using a tool that's towards your goal, and then you just let it cook.”
Quotes lightly edited to remove filler words.
From the conversation
This explainer is drawn from these episodes — each carries its full transcript.
