Forty-five AIs working together discovered 266 software vulnerabilities, while the same AIs working alone found only 21. We believe this comparison matters more than simply “running several AIs”: AI’s capability ceiling is shifting from individual intelligence to whether it can be organized into teams and obtain real feedback.
What this is
The article argues that multi-agent systems—multiple AIs dividing tasks, sharing information, and working together—are not equivalent to simply copying a model several times. If multiple AIs merely inspect the same material, the gains are limited. Extra value emerges only when they can access information a single AI could not obtain on its own, such as running tests, reviewing rendered webpage screenshots, or using tools to verify facts.
The article cites a self-correction experiment in which an AI reviewing and revising only its own answers saw accuracy fall from about 26% to an even lower level. When the AI generated a webpage and was then given both the actual rendered screenshot and its description, the pass rate rose to about 52%. This suggests that “thinking a few more steps” is less effective than “looking once at the real result.” The former may simply repeat deductions along an existing line of reasoning, while the latter can expose discrepancies between code and expectations.
Industry view
Our view is that AI competition is shifting from single-model capability toward workflow design. Differences between models still matter, but actual output often depends on whether task decomposition, context sharing—ensuring later participants understand the background to earlier work—tool use, and acceptance testing are reliable.
The risks are equally clear. More steps do not necessarily produce better results: AI may stop early or settle for a superficial answer. Multiple AIs can also significantly increase inference costs. Organizational design does not improve indefinitely with scale, either. Sharing every record can create information overload, while isolated records require additional communication mechanisms. Collaborative AI may also amplify errors across agents, making independent verification and a clearly accountable final owner essential.
Impact on regular people
For enterprise IT: When enterprises procure AI, they should not compare only model parameters and the quality of a single answer. They must also calculate the inference, verification, and failure costs of multi-agent systems on real tasks. A single model with a clear workflow and accessible tools may be more cost-effective than an expensive collaborative system.
For individual professionals: Personal AI use also benefits from an “independent first, external check second” approach. After an AI produces a result, run tests, check original sources, or inspect the actual page rather than repeatedly asking it to debate with itself.
For the consumer market: As AI products begin emphasizing team collaboration, pricing may shift from charging per call toward charging per completed task. Consumers may gain stronger workflows but also bear more process-related costs. The standard should return to verifiable outcomes, not the number of AIs opened simultaneously.