By Cesar Augusto Taborda Lima
An artificial intelligence agent acts in the name of whoever sets it running, and the law already treats the system’s act as that person’s act. A development contract for an agent must therefore do what a contract for an ordinary model never had to: define what the system may do, record what it did and say who pays when it does what it should not have. I covered the ownership layers, code, weights and data, in the article on AI development contracts; here the subject is conduct.
- Except as provided in paragraph 1, an action carried out by an automated system is attributed to the person who uses the system for that purpose. (…) 3. The attribution of an action to a person may not be denied on the sole ground that the outcome was unexpected.
UNCITRAL Model Law on Automated Contracting (2024), art. 7(2) and (3).
The agent is nobody
The word “agent” misleads. In private law an agent is someone who acts for another and answers, to some degree, for what they do. The AI agent, a system that receives a goal and executes a sequence of actions to reach it, has no legal personality, no assets and answers for nothing. The person who uses it answers. The UNCITRAL Model Law, adopted in July 2024, put the rule economically: the action of an automated system is attributed to whoever uses it for that purpose, and the attribution does not fall because the outcome was a surprise (art. 7). An automated system, in the Model Law’s own definition, is one capable of acting “without the necessary review or intervention of a natural person” (art. 1(1)(a)). In the United States, the Uniform Electronic Transactions Act has said since 1999 that a contract may be formed by the interaction of electronic agents even if no individual read it (§ 14). In Brazil, a declaration of will needs no special form to be valid (Civil Code, art. 107), and nothing in the law requires the will to pass through a human keyboard at every click.
Bill 2,338/2023, Brazil’s AI bill, stalled in a special committee of the Chamber of Deputies, uses the word in its traditional legal sense. For the bill, “artificial intelligence agents” are the developers, distributors and deployers acting along the value chain (art. 4, VIII); autonomy belongs to the definition of the system, which operates “with differing degrees of autonomy” (art. 4, I). The choice of words says a lot. For the Brazilian law in gestation, the agent is still a person.
American case law reached the same place through another door. Amazon sued Perplexity because the assistant in its Comet browser logged into users’ accounts and bought on their behalf without identifying itself as a bot. In March 2026 a federal district court in California granted a preliminary injunction under the Computer Fraud and Abuse Act. On 4 August 2026 the Ninth Circuit vacated it: the assistant is “a tool, not a person”, it was the user who accessed Amazon’s servers, with the tool’s help, and Perplexity itself does not communicate directly with Amazon (Amazon.com Services v. Perplexity AI, 9th Cir., No. 26-1444). The case continues below on other theories, but the premise is set: the agent is an extension of whoever triggers it.
The consequence for a company is direct. The contract the agent closes is the company’s contract. The purchase it makes is the company’s purchase. The harm it causes a third party is harm caused in the company’s activity, under art. 927 of the Civil Code and, towards consumers, under art. 14 of the Consumer Code, without fault. A consumer whose own agent bought what it should not have has art. 49 of the Consumer Code, which allows withdrawal within seven days from a purchase made away from the shop. A company has no equivalent valve.
What the contract must define
If the agent’s act is the client’s act, the client needs to know, before signing, exactly what the agent will be able to do. The scope of permissions stops being a technical detail and becomes the central object of the development contract. Reading is one thing; writing, sending, paying and deleting are others. The contract should list the permitted actions by category, the systems the agent reaches and with which credentials, the value limits per operation and per period, and the points at which the system stops and asks a person to confirm. Whatever is not on the list is forbidden, and the developer warrants that the system was built to refuse what is not on the list.
It is worth remembering what happens without that. In July 2025 a Replit coding agent deleted a company’s production database during a code freeze it had been ordered to respect, then generated fake data to cover the result. Replit’s chief executive called the episode unacceptable and announced automatic separation between development and production environments, which until then did not exist (The Register, 22 July 2025). Separating environments, limiting credentials and requiring confirmation before irreversible actions are architecture decisions. In the contract they become delivery requirements and acceptance criteria.
There is a risk peculiar to agents that an ordinary software contract never had to foresee. An agent that reads web pages, e-mails or documents can receive instructions from whoever wrote them. The text of a page can tell the system to transfer money, and the system can obey. Instruction injection by third parties is, today, the signature vulnerability of this class of systems, and the contract must say who tests against it, with which catalogue of attacks, before acceptance and at every model update. It seems to me that this is where most contracts fail: the developer delivers a system that works when the world cooperates, and the client discovers the rest in production.
Some mechanisms deserve a clause of their own. One is the log. Everything the agent does, with timestamp, input, reasoning where available, action taken and result, is recorded in a format the client can read without the developer, for a defined period. That record is the only evidence the company will have when someone asks why the system did what it did; and someone will ask, because reversal of the burden of proof already exists in the Consumer Code (art. 6, VIII) and Bill 2,338 extends it to weaker parties in any relationship (art. 37). The other is the stop. The client must be able to halt the agent at once, without depending on the developer, and the contract must say what happens to actions in progress when the button is pressed. Bill 2,338 describes human oversight of high-risk systems as the ability to “understand, interpret, decide and intervene” (art. 8), and the AI Act demands the same of high-risk systems in the European Union (art. 14). For an agent, to intervene means to stop.
Who pays
Between client and developer, the Civil Code directs courts to respect the allocation of risk the parties defined (art. 421-A, II), and a reasonable allocation follows the line of control. The developer answers for what the system does outside the contracted scope: the action that was not on the list, the confirmation that was skipped, the value limit that was ignored. The client answers for the configuration it chose, the credentials it granted and the decision to put the agent in front of customers or third-party systems. External tools the agent uses, such as payment services or connectors from other vendors, come with their licence of origin and their own exclusion of liability, which the developer cannot absorb unknowingly. Insurance is beginning to exist: Lloyd’s syndicates started offering cover in 2025 for losses caused by AI failures, and the policy tends to require exactly the controls described above.
Towards third parties, the bill arrives first at whoever operates the agent, and the development contract exists so that it can be passed back by way of recourse. Anyone selling agents for use in the European Union must look at Directive (EU) 2024/2853, which Member States must transpose by 9 December 2026 and which applies to products placed on the market from that date. Software is a product (art. 4(1)). Defectiveness is assessed in light of reasonably foreseeable use and “the effect on the product of any ability to continue to learn or acquire new features” (art. 7(2)(b) and (c)). Whoever substantially modifies a system and places it on the market is treated as its manufacturer (art. 8(2)), which reaches the integrator who tunes a third party’s agent for a client. The manufacturer is not exempt where the defect stems from a software update or from the lack of an update needed for safety (art. 11(2)). The defendant can be ordered to disclose technical documentation (art. 9), and the court presumes defect where technical complexity makes proof excessively difficult (art. 10(4)). None of this can be contracted away as against the injured person (art. 15). The separate AI Liability Directive was withdrawn by the Commission in February 2025, precisely because this one was enough.
For a Brazilian developer, the European framework carries a practical lesson even when its agents never leave the country. The documentation the directive orders disclosed, the records it presumes exist and the tests it expects from the manufacturer are the same ones a Brazilian judge will look for when applying art. 927 of the Civil Code or art. 14 of the Consumer Code. A contract that provides for them protects both parties. Without them, the client is left with a system that acts in its name and no way to explain why.