THE POINT / WHAT CHANGES FOR YOU

From news to practice.

Before granting autonomy to an agent, define objectives, tools, and permissions. Plan human checks for important actions and limits to automatic work.

An agent does more than just respond

An assistant can suggest a reply; an agent can choose steps and use tools to achieve a goal. It can search documents, query an archive or prepare an action in an external system. Autonomy, however, is a design choice, not a quality to be increased without limits. The primary task is to establish what the system should achieve and which decisions should remain with people.

Start from the problem, before the technology

A useful project starts with a verifiable outcome: for example, classifying support requests and preparing a draft, without sending it. Real examples, correctness criteria and a process owner are needed. If a sequence of rules or a single request to a model already solves the task, an autonomous agent may add complexity without sufficient advantage. This is the principle of simplicity also recalled by Anthropic in the guide Building effective agents.

Four architectures to know

In the sequential flow, predetermined steps lead from input to result. In routing, a request is directed to the appropriate path or specialist. In the single agent, a model selects tools and steps within defined limits. In the system with multiple agents, distinct roles collaborate or hand off work. No structure is better in absolute terms: the additional coordination must be justified by measurable results, because it can increase costs, time and possibility of error.

How to choose the model underlying the agent

OpenAI's GPT families, Anthropic's Claude, Google's Gemini and models from other providers are possible components, not complete agents. To choose you must test the specific version on your own tasks: understanding instructions, correct use of tools, document management, timing and cost per completed activity. A cheaper model per request may be less convenient if it requires more attempts and corrections. Performance should be measured on the process, not inferred from the brand.

Few tools, well described and with limited permissions

Each tool must have a recognisable purpose, controllable parameters and verifiable results. Reading and modifying a customer record must be separate authorisations. An agent preparing a draft does not need, for this reason only, permission to send it. The application must enforce limits even when the model produces a wrong request: writing a prohibition in the prompt does not replace technical control.

Data, memory and work state

An agent must know which information is valid and to which activity it belongs. It is useful to distinguish reference documents, conversation memory and operational state: what has been done, what is missing and what approval has been received. A system that resumes an interrupted execution must avoid repeating actions already completed. Frameworks like LangGraph dedicate tools to state management, interruptions and process resumption.

The dangers: malicious instructions, exposed data and wrong actions

A page or document may contain instructions attempting to mislead the agent: this is the risk of prompt injection. Other issues concern excessive access, unnecessary information sharing, errors propagated between steps and sequences that continue without reaching the goal. OWASP dedicates a specific guide to threats of agentic systems. Risks must be evaluated case by case: an agent that reads an archive and one that modifies orders have different potential consequences.

Controls must be designed alongside autonomy

For an initial release tools, duration, number of steps and expenditure can be limited, recording events necessary to reconstruct the work. Relevant actions must have a human approval point, with visible recipient and content. Stop conditions and a pathway to transfer the case to a person are also needed. These precautions reduce exposure to errors but alone do not guarantee system security.

An example: the agent that assists an editorial office

The goal could be to prepare a review from a list of authorised sources. The system collects documents, keeps URLs and dates, proposes a summary and flags unconfirmed information. The editor checks sources and attributions; publication requires a separate decision. To measure the project, omissions, errors and review time are compared with previous work. The example is a design outline, not a service already implemented by IAnews24.

Test before extending the scope

A useful test includes normal requests, missing data, ambiguous instructions, and manipulation attempts. Observations include task completion, errors, human interventions, time, and cost. After a model or tool change, important cases must be repeated. The most reliable path is a circumscribed project with verifiable results, followed by gradual extension. The value of an agent is measured by the work it completes correctly, not by the number of actions it performs.

Ten companies to watch in the global AI agents market ↗

Sources and documents

Document date or reference period: Guides and technical documentation consulted on 23 October 2026.

Text prepared with AI assistance based on the indicated sources. Interpretive evaluations are distinct from data and statements from sources.

Report an update or correction →