AI agents are expected to research the web, run code and handle tasks that require several steps. But the training pipelines used to develop those capabilities are becoming increasingly complex. A team from TikTok parent Bytedance and several US universities argues that developers may be able to get there with three steps.
Why training pipelines are getting harder to follow
Many developers now stack multiple training phases one after another. As these pipelines grow more nested, it becomes difficult to determine which phase is responsible for a model’s abilities.
That uncertainty has practical costs. If researchers cannot separate the contribution of each stage, it is harder to understand why a system works and to reproduce the process. The paper says reproducing this work is becoming expensive.
The issue is not simply the number of steps. Each additional phase can make the overall method harder to inspect. When several stages are combined, a successful result does not by itself reveal which parts were necessary.
A case for using three steps
The paper’s authors aim to show that three steps are enough. Their argument challenges the assumption that more elaborate training sequences are needed to produce agents that can take on demanding tasks.
A shorter process could make it easier to see how training contributes to a model’s behavior. With fewer phases to account for, researchers may have a clearer path to examining the method and repeating it. That is the central practical appeal of the proposal.
The source describes the work as being backed by a competitive open model. The goal, as presented, is to pair a simpler training approach with a model that can compete, rather than treating fewer steps as an end in themselves.
What simpler training could change
If the three-step approach holds up, it could give researchers a more manageable way to study how AI agents acquire capabilities. A pipeline with fewer nested phases would make it easier to ask which parts are doing useful work, and whether the same outcome can be reproduced without a long chain of stages.
That could matter for developers trying to build systems that research the web, execute code or work through multi-step problems. The paper’s concern is that complexity makes those methods difficult to evaluate and costly to reproduce. A more compact process could make the underlying choices easier to understand.
Still, the article’s description does not provide the details of the three stages or report specific evaluation results. It therefore supports the broader claim that the team is proposing a simpler route, but not a detailed comparison of how the approach performs against other training pipelines.
The question is what each phase contributes
The broader debate raised by the paper is about how much training complexity is actually needed. As developers add phases, it becomes more important to know which ones improve an agent and which ones simply make the process harder to trace.
Bytedance’s team and its university collaborators are putting forward a three-step alternative and a competitive open model as evidence for that case. The proposal offers a clear test for future work: whether a shorter training pipeline can support capable AI agents while making the process easier to understand and reproduce.