149 runs, 275 published segments, only 2 actually revised — these are the internal numbers a Chinese financial research Agent team disclosed on Juejin. Behind them is a counter-intuitive finding: the more capable the Agent, the more the final answer feels like several research reports mechanically stitched together. Their fix was not a bigger model. They rewrote the writing goal and the context management.
What this is
"Progressive Answer" means an AI Agent pushes a new segment to the user each time it gets a fresh batch of evidence, instead of waiting until the entire research job is done to deliver. The intent is to reduce perceived waiting time, but in engineering practice it tends to split the answer into multiple "mini reports" — each carrying its own background, summary, and risk disclaimer. Combined, the user is left to figure out which segment is the real final conclusion.
Their core refactor had two steps. First, they changed the Agent's writing goal from "each segment stands on its own" to "always writing the same article." Second, they let the model clearly see "what it has already written" before each continuation, instead of only seeing the conversation history (this falls under Context Engineering — how to feed the right content to the model). They also shifted the evaluation metric from "when does the first segment come out" to "first useful answer," emphasizing that what should be delivered early is the answer, not the progress.
Industry view
Supporters call this an honest engineering reflection. Many teams blame a poor Agent experience on the model not being strong enough; the article argues the problem often lies in the delivery mechanism. At minimum, this confirms one thing: as model capabilities converge, the competitive focus is shifting from "parameters" to "process design."
But there are dissenting voices. First, this kind of refactor is highly vertical-dependent — financial research has a clear argumentative spine you can "keep writing around," whereas open chat or creative writing may not apply. Second, the article inadvertently exposes the awkward state of today's Agent frameworks: context management matters more for the final user experience than the model itself, yet there is no standard solution, and every vendor builds its own. Third, the article ends abruptly — core technical details are not fully disclosed — making replication hard for outsiders. In other words, whether this method is truly "general" still needs more evidence.
Impact on regular people
For enterprise IT: When selecting internal Agent projects, the evaluation focus may need to shift from "is it wired up to a top-tier large model" to "how is context managed and how is the answer organized" — the latter has a bigger impact on the end-user experience.
For individual professionals: If you often use AI to write weekly reports or research material and find the output feels like several reports stitched together, it may not be that the AI is bad — the tool configuration or prompts may need tuning.
For the consumer market: This won't immediately flow through to consumer products, but it reminds us of one thing: the gap to AI assistants that "feel human" is not IQ — it is structure.