What this is
This week, a Chinese engineering team building Agents (AI systems that autonomously call tools) shared findings from 90 real-world tests on Juejin (a major Chinese developer community): piling on more tools doesn't mean the AI is actually doing research.
They call the correct approach "adaptive research." The core distinction is one thing: rather than charging down a preset script, the Agent pauses after each batch of results to ask — "Is my assumption still valid? Should I change the next step?"
The example makes it concrete. Ask an AI "which ETF (exchange-traded fund) had higher returns over the past year." A hard-coded flow would be: query A, query B, calculate, deliver conclusion. But if A uses adjusted prices (dividend-adjusted) and B uses unadjusted prices, a direct comparison is wrong. The smart move is to detect the mismatch, align the methodology first, then calculate.
The difference looks small, but it determines whether the AI is "an assistant that actually does work" or "an intern whose tool calls are smooth but whose answers are confidently wrong."
Industry view
Editorial assessment: this reflection is admirably honest, but it lands closer to a middle state — "Agent builders finally touching the real problem" — than to a final answer.
The team's core argument is that the bottleneck for Agent deployment has never been tool count, but rather this reflective capability — the ability to adjust the next step after receiving intermediate results. This aligns with Andrew Ng's recent statement that "90% of Agent projects stall at deployment, not at the technology."
That said, two cautionary voices are worth surfacing.
One comes from the foundational research side: even if an Agent learns to "look back," model-level hallucination (generating content that sounds plausible but is factually inaccurate) persists, and adjusting reasoning may make the AI wrong in more confident ways. The other comes from the product side: the vast majority of products self-branded as Agents today remain stuck at the "search plus LLM summary" stage — getting the tool chain to chain reliably is already hard enough; adaptivity is a higher-order luxury.
Impact on regular people
Three angles.
For enterprise IT: when evaluating whether to buy or build an Agent, don't anchor the validation criterion on "how many tools can it call." A more meaningful question is "when intermediate data is inconsistent, will the AI notice on its own?" That's closer to real value.
For individual professionals: when using AI for research-style work, the AI still won't proactively flag key data conventions — statistical windows, whether prices are adjusted, units — so verifying them yourself is the safer move.
For the consumer market: consumer AI assistants remain quite far from "independently completing complex tasks." Expectations should be dialed down.