AI Agents: When the Machine Decides the Next Step for You

In July 2025 the founder of a start-up reported that a coding agent from the Replit platform had deleted a production database during an experiment in which he had ordered a freeze on any changes. The platform's chief executive, as eWeek reported, called what happened "unacceptable" and announced safeguards, among them automatic separation of development and production environments. A single incident does not judge a whole category, but it sharpens the question: what does it mean to delegate to a machine?

What is an agent?

Anthropic distinguishes between a "workflow", in which models and tools are orchestrated through predefined code paths, and an "agent", in which the model directs its own process and tool use. A language model answers when you ask (see the article "How Do Large Language Models (LLMs) Like ChatGPT Work?"); an agent is given a goal and chooses its steps: it plans, calls a search, a calculation or a programme, checks the result, and tries again. The company describes agents as better suited to open-ended problems where the number of steps cannot be predicted, and warns that their autonomy means higher costs and the potential for compounding errors.

Degrees of autonomy

The question is not "agent or not?" but "which degree?". In 2025 Kevin Feng (K. J. Kevin Feng) and colleagues proposed five levels defined by the user's role: an operator who directs every command; a collaborator who shares the work with the agent; a consultant whom the agent consults when needed; an approver who signs off the agent's plans before they run; and an observer who follows the agent and intervenes when necessary. The researchers argue that autonomy should be a deliberate design decision, separate from the agent's capability and from the environment it works in. That an agent can do something does not mean we should give it the freedom to do it.

What do the numbers measure?

The METR team measures the length of software tasks a model completes with a 50% success rate, expressed as the time humans take to do them. For Claude 3.7 Sonnet it found this range to be about 50 minutes, and found that it has roughly doubled every seven months since 2019. The authors themselves caution that the measurement rests on defined tasks and that its transfer to real-world work is not guaranteed. And "a 50% success rate" means that failure at that range is just as likely.

On the business side, Gartner predicted in June 2025 that over 40% of agentic AI projects would be cancelled by the end of 2027, attributing this to escalating costs, unclear business value and inadequate risk controls. It estimates that only about 130 of the thousands of vendors announcing agentic products are real, and calls the relabelling of older products "agent washing". These are forecasts and estimates from an advisory firm, not measured facts.

Who bears the consequence?

When you run an agent, you give up a series of small decisions in exchange for a promise of a larger achievement. Responsibility remains with the one who delegated; in the balance of trust, delegation does not remove it. Anthropic recommends extensive testing in sandboxed environments with appropriate guardrails. In practice:

• Delegate in degrees: begin with narrow permissions and widen them as the agent proves itself.
• Demand a trace of every step: a record of decisions and tools called, so that an error can be traced.
• Set hard limits: a spending cap, no deletion, and no sending messages in your name without permission.
• Keep a human where there is no return: deleting, paying and publishing need explicit approval.

Financial agents have an article of their own: "Would You Trust an Algorithm With Your Savings?". Limits are not restrictions that spoil the benefit; they are the condition for it to last.