Agents, Automation and the Future of Work

An agent is a model that does not just answer but acts: it plans a sequence of steps, calls tools (search, code execution, a browser, a company's APIs), observes the results and keeps going until a goal is met or it gives up. Chatbots sell answers; agents sell completed work. That shift, which the industry pushed hard from late 2024, changes the business question from "how much is a better answer worth?" to "how much of a job's labour cost can software take?", a pool of money far larger than any software budget. It also changes the risk. A wrong answer is read by a human; a wrong action is executed. Agents in 2025–26 became genuinely useful in a few domains, above all software engineering and customer support, where the work is digital, the steps are checkable and the cost of an error is bounded. Elsewhere they remained unreliable over long tasks, expensive to supervise and hard to integrate. The labour-market evidence is mixed and early: large-sample studies through 2025 found no economy-wide disruption, but some found falling employment for young workers in the most exposed occupations, and several large companies attributed headcount cuts to AI. The subject is therefore an answer to four questions: what can an agent reliably complete without a human, what does a completed task cost compared with a person, who carries the risk when the agent errs, and what happens to the jobs, firms and service industries whose work it takes over? This section explains how agents are built, why reliability compounds, the products and standards of 2024–26, the "services-as-software" thesis, the economics of a deployed agent, the labour-market evidence and its limits, landmark cases, pitfalls and expert framing. Pricing models for agents (per resolution, per outcome) are covered in 11; evaluation and safety of agents in 14.