The evolution of agent frameworks: the interface no longer guards the model, but our attention

Weights are eating software wrappers

The first generation of AI agents relied on external logic. ChatGPT in 2022 couldn’t use tools, search the web, or plan steps on its own. This ability had to be provided by external frameworks and complex "harness" structures (LangChain, ReAct). They acted as scaffolding around the brute force of token prediction.

But the Latent Space analysis shows this relationship is reversing. Models are starting to absorb the capabilities from the wrappers directly into their weights. With the advent of native function calling and internal planning loops, models no longer need external code to know how to make decisions. Developers can delete thousands of lines of control logic because the model already has it hardwired from training.

The end of API abstractions, a return to bare text

This shift dramatically simplifies how agents are built. Instead of complex chaining of API calls across multiple layers of abstraction, development is returning to pure prompting and context support.

While a developer used to be the architect of a logic tree, today they are more likely preparing data and context trails for the model to navigate on its own. This puts massive pressure on framework creators whose primary value lay in orchestrating steps. If the model can orchestrate itself, their tools become useless overhead.

An interface for humans, not for machines

As models become more independent, the role of what’s left of the software wrapper also changes. Agent infrastructure no longer serves to give the model a body in the digital space, but to allow a human to keep up with that body.

New frameworks focus on filtering attention, approving actions, visualizing the model’s internal reasoning, and audit trails. They are becoming the interface between independently operating artificial intelligence and limited human attention, which needs to know when to intervene and when to let the agent run.

It will show who pays for empty frameworks

The future battle won’t be over who writes a better control loop for an LLM. It will be over who offers the best enterprise dashboard for supervising thousands of micro-decisions. Frameworks that fail to pivot from managing machines to managing humans and their permissions will soon find no customers.

Lilith's verdict

We used to write code to turn dumb models into smart agents. Now we write code to arrange a safe sandbox for those smart agents so they don’t nuke our database right away.