AT A GLANCE
An agent can return something in minutes and still leave a person with hours of checking. Measure the cost of a completed job, including review and rework, and agree who can approve spending.
The InfoWorld feature asks how we’ll check software when agents write code we struggle to understand. The same question applies to the agent you put in charge of comparing supplier quotes. Who decides when it is ready to place the order?
Ethan Mollick’s Claude Fable 5.2 experiment produced 204 interactive worlds in 20 hours. Its agents then wrote an 87,500-word rulebook for quality control and the pace dropped to about a dozen worlds a day. Output picked up again after Mollick intervened and had agents review the process. Mollick’s account (opens a new tab)
In OpenAI’s 2025 evaluation of 220 professional tasks, experts judged GPT-5’s output better than or on par with human work in 39% of comparisons. The model could produce a draft much faster. In the authors’ analysis, including review and redoing weaker work reduced the estimated savings to about 11% in time and 15% in cost in their conservative scenario. OpenAI’s study (opens a new tab) and the evaluation paper (opens a new tab)
At Ignite, Microsoft put more work inside Copilot and previewed spending controls for agents. That connects to the strategy I wrote about in #25 on 27 August 2023. The agent comparing quotes today will soon be asked to place the order. Microsoft Ignite 2026 coverage
Run the trial on a real task. Measure an outcome a person is willing to sign off, including repeat runs, review and corrections. Agree who can raise the spending limit, who can stop the agent and which decisions still need approval.
THE SOURCE / FURTHER READING
Want the full feature?
You’ve read the TOP STORIES interpretation. Continue to the publisher for the source feature or detailed guidance.
Read the original feature at InfoWorld (opens a new tab)