If you hear “agent” and think of the sales rep, you’re not wrong. Italian law sees it the same way: under the civil code, an agent takes on a standing job of promoting contracts on someone else’s behalf, in a given area.
We mean something else here, though the two have a little in common. An AI agent is an artificial intelligence you hand a whole job to. You don’t tell a rep every step. You give them the job and they get on with it. That is where the likeness ends. A rep answers for what they do. An agent doesn’t.
How it differs from a chat
A chat answers a question. You ask, it replies, it stops. The next step is yours.
An agent takes a job and works out for itself how to finish it. Anthropic, the US company behind Claude, an AI along the lines of ChatGPT, puts it this way: the model directs its own process and the tools it uses. Those tools might be your email or your business software. OpenAI, the company behind ChatGPT, is blunt about it: a simple chatbot is not an agent.
An example, made up
This is an example, not a real case. A customer dashes off an email: “Same boxes as last time but smaller, need them by Friday.”
The agent reads it. It looks up that customer’s last order in your system. It fills in a works order with what it finds. It spots that the size is missing and drafts the question: “Good morning, what size would you like for Friday’s boxes?”
The order and the question land in front of someone in your office. They read them, change anything that needs changing and send with one click. The email to the customer comes from a person. The customer’s answer stays on the order.
What it can do, and where it goes wrong
OpenAI recommends agents for work that means interpreting everyday language and making sense of documents. We’d say an email written in a hurry fits that.
Stanford’s AI Index 2026 says agents still fail roughly one attempt in three on structured tests. On OSWorld, a test of computer tasks, the best agent in the report reaches 66.3%, against about 72% for a group of computer science students. In March 2026 OpenAI claimed 75% for one of its models, on a revised version of the test: that figure is OpenAI’s, not Stanford’s. In a small software company simulated by researchers, the best agent finished 30% of tasks on its own. The figure goes back to May 2025 (latest version of the study: September 2025).
The mistakes are down to earth. In 2025 Anthropic had Claude run a small shop in its office. It told customers to pay into an account it had made up, and it sold at a loss. Anthropic wrote that it would not hire it. In the second phase, with newer models and written procedures, loss-making weeks largely disappeared. Even so, the agents were drawing up a contract that fell foul of a US law, until a member of staff stopped them.
One risk has to do with email itself. A message can hide instructions for the agent in white text you can’t see. This is called prompt injection. The UK’s National Cyber Security Centre says it may never be fully mitigated. In a prize contest among experts, published by Anthropic in August 2026, attacks worked in 0.08% of attempts against Claude Opus 5, 0.58% against the best model from another company (Kimi K3) and 8.09% against the weakest (DeepSeek V4 Flash). The 11 tests covered tools, code and computer use: it isn’t a figure about email. Anthropic’s models ran without the extra protections of its products. Those are Anthropic’s figures, and Opus 5 is its own model. Low does not mean zero.
Why a person decides
Anthropic says people should stay in control, above all before decisions that carry real weight. OpenAI advises having a person in front of sensitive or irreversible actions, such as cancelling an order or making a payment, until confidence in the agent grows.
A click is not enough if nobody reads what they are approving. The EU’s AI Act has a name for this: automation bias, which means leaning too heavily on what comes out of the machine. Article 14 is about high-risk systems. We think the temptation comes up in front of an order too.
How you keep it in check
There are four things to ask of anyone who offers you an agent. The agent prepares and a person sends: nothing goes out to a customer or supplier on its own. It has written limits: how many times it retries, when it stops and hands over. Everything leaves a trail: what it read, what it proposed, who sent it. And you count: errors a month before, errors a month after.
A hunch won’t do. In 2025 the research group METR measured 16 experienced software developers. With AI they took 19% longer, and they believed they had been 20% faster. METR itself writes that developers are probably faster now than they were then, but that its newer data is unreliable. The gap between impression and measurement still shows why you count.
Our job is to bring agents into businesses that make things, on these terms: the starting number gets written down first, and a person makes the decision.
Sources
- Codice civile (Italian civil code), art. 1742, text on Brocardi.it. www.brocardi.it
- Erik S. and Barry Zhang, Building effective agents, Anthropic, 19 December 2024. www.anthropic.com
- OpenAI, A practical guide to building agents, OpenAI, April 2025 (date of the PDF file). cdn.openai.com
- Stanford HAI, AI Index Report 2026, chapter Technical Performance, Stanford University, 2026. hai.stanford.edu
- Stanford HAI, AI Index Report 2026, chapter Technical Performance (PDF), Stanford University, 2026. hai.stanford.edu
- OpenAI, Introducing GPT-5.4, OpenAI, 5 March 2026. openai.com
- Frank F. Xu et al., TheAgentCompany: Benchmarking LLM Agents on Consequential Real World Tasks, arXiv, version 3, 10 September 2025. arxiv.org
- Anthropic, Project Vend: Can Claude run a small shop? (And why does that matter?), Anthropic, 27 June 2025. www.anthropic.com
- Anthropic, Project Vend: Phase two, Anthropic, 18 December 2025. www.anthropic.com
- Anthropic, Mitigating the risk of prompt injections in browser use, Anthropic, 24 November 2025. www.anthropic.com
- National Cyber Security Centre, Prompt injection is not SQL injection (it may be worse), NCSC (United Kingdom), 8 December 2025. www.ncsc.gov.uk
- Anthropic, System Card: Claude Opus 5, Anthropic, 24 July 2026, updated 19 August 2026, section 5.2.2.1. www-cdn.anthropic.com
- Anthropic, Our framework for developing safe and trustworthy agents, Anthropic, 4 August 2025. www.anthropic.com
- Regulation (EU) 2024/1689 (AI Act), Article 14, Italian text, AI Act Service Desk of the European Commission. ai-act-service-desk.ec.europa.eu
- Regulation (EU) 2024/1689 (AI Act), Article 14, English text, AI Act Service Desk of the European Commission. ai-act-service-desk.ec.europa.eu
- Joel Becker, Nate Rush, Beth Barnes, David Rein, Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, METR, 10 July 2025. metr.org
- METR, We are Changing our Developer Productivity Experiment Design, METR, 24 February 2026. metr.org