FT reporters say AI agents speed up work but add heavy checking

In short: Financial Times data reporters say AI “agents” can produce work fast, but they also create a lot more to double-check.

What's going on

Financial Times data reporters John Burn-Murdoch and Clara Murray shared how “AI agents” are changing their work. An AI agent is a tool that can carry out a longer task on its own, like a junior assistant who can write code and fetch information while you wait.

Murray tested OpenAI’s Codex, an agent tool for coding, on two projects. One pulled product prices from websites in multiple languages. The other built an interactive calculator using millions of data points. She said she could get prototypes quickly and even run many agents at once, but she did not end up working less.

As the agents produced large amounts of code and data, Murray felt she had to stay glued to her screen to supervise. Her laptop even crashed under the volume. By the end of the week, she found major errors, including an agent that “guessed” prices by picking the biggest number on a web page, and other mix-ups such as product variants in different colours.

Burn-Murdoch described a similar problem. Agents can return lots of research and calculations, but that output needs “voluminous verification” before publication. He said he uses a process he calls an “adversarial audit”, where he asks another AI system to look for mistakes (like having a second person check your homework), then he manually checks the most important parts.

What to watch

This points to a broader shift in many jobs. AI agents may move people from doing tasks to managing and checking them. For journalism and any work where mistakes matter, the key question is whether teams can build reliable checking steps fast enough to safely use agent output at scale.

Source: Financial Times