What is an AI agent, and how does it differ from an assistant?

The agent holds the keys to e-mail, the database and payments; every call it makes is checked before execution against a rule signed by a person, and the decision is recorded so that it cannot be altered afterwards.01 · WHAT AN AGENT ISA model with the keyse-mail₽paymentdatabasecodeThe program itself decides which tool to calland calls it: a payment, a letter, a record.02 · WHAT IS JUDGEDEvery call, before it runsRULEpay()allowedrefusal with a reasonEach action is judged by a rule a person signed;unproven, it returns as a refusal with a reason.03 · WHAT REMAINSThe rule and the recordsigned by a personexternal timestampreceipt chainEach decision is recorded so that it cannot bealtered afterwards; an outsider can check it too.The agent holds the keys to e-mail, the database and payments; every call it makes is checked before execution against a rule signed by a person, and the decision is recorded so that it cannot be altered afterwards.01 · WHAT AN AGENT ISA model with the keyse-mail₽paymentdatabasecodeThe program itself decides which tool to calland calls it: a payment, a letter, a record.02 · WHAT IS JUDGEDEvery call, before it runsRULEpay()allowedrefusal with a reasonEach action is judged by a rule a person signed;unproven, it returns as a refusal with a reason.03 · WHAT REMAINSThe rule and the recordsigned by a personexternal timestampreceipt chainEach decision is recorded so that it cannot bealtered afterwards; an outsider can check it too.
The agent holds the keys to e-mail, the database and payments; every call it makes is checked before execution against a rule signed by a person, and the decision is recorded so that it cannot be altered afterwards.

An agent built on artificial intelligence (AI) is a program based on a large language model that does more than generate text: it carries out operations in external information systems on its own. Unlike a chat assistant, an agent is given the authority to call application programming interfaces (API), change records in databases, run code and start transactions on behalf of an organization.

A chat assistant gives a person an answer, while an AI agent sends a command straight to the execution environment: a database, a cloud server, a payment gateway or a file store. Such connections already work at industrial scale: according to Anthropic, as of 9 December 2025 there were more than 10,000 active public servers of the Model Context Protocol (MCP), and its software development kits (SDK) for Python and TypeScript were downloaded more than 97 million times a month. If an agent is handed full-permission keys with no check in between, a mistake of a probabilistic model immediately becomes a deleted production volume or a wiped file system. Manual approval of every operation, on which 37 of 70 public descriptions of agent deployments in Russia in 2025–2026 rely, stops working under a stream of actions. That is why the parameters of every call are checked by a solver against a machine-readable contract, which separates what the model proposes from what is admitted for execution.

What an AI agent is in simple terms

In practice, an AI agent differs from earlier programs in its right to act on its own. It receives a goal, breaks it down into intermediate steps, chooses software tools and runs a chain of calls until it gets a result. A person sets the overall task but does not type each command.

Architecturally, agentic AI is a system of four connected parts.

  1. The language model (planning core): turns the task into a sequence of machine operations and interprets the responses of external programs.
  2. Memory and state: keeps the context of a multi-step task, the history of calls already made and data taken from corporate reference systems.
  3. The tool catalog (the interfaces of action): modules connected through standard buses, in particular through the MCP protocol, for working with files, querying databases, sending messages or managing cloud infrastructure.
  4. The execution environment: a server, a workstation or a sandbox where commands and scripts run under the credentials that were issued.

In engineering practice this class of systems is called agentic AI. Classic robotic process automation (RPA) scripts follow a fixed route and stop at any change in the interface. An agent that runs into an error looks for a workaround on its own to finish the task. This ability to improvise while holding credentials is what creates a new class of information security threats.

How an AI agent differs from a chatbot and an assistant

The difference between an AI agent and an AI assistant is a direct path to the infrastructure. An assistant finishes its work by showing text on the screen of the user, and an agent finishes it by changing the state of an external system. Between the advice of an assistant and the production environment there is always a person who reads the answer and decides whether to run the proposed code. In an agent, this buffer is removed for the sake of speed.

FeatureChatbot (scripted)AI assistant (conversational)AI agent (autonomous executor)
Result of the workAn answer from a pre-written decision treeGenerated text, a reference answer, a draft of codeA completed operation: a changed database, a sent API call, a script that ran
Degree of autonomyNone: it moves strictly along the branches of the scriptOne step: a human’s request, a text answerMany steps: planning, calling tools, reading errors, working around obstacles
Access to infrastructureReads a reference list or files a request in one formReads files uploaded to the chat and search resultsCalls APIs directly, has access to a terminal, the file system and cloud resources
Cost of a model errorAn unhappy user, a hand-off to an operatorAn inaccuracy in the answer, which an employee correctsDeleted working volumes, files and databases
ControlThe script tree is checked when it is set upA person proofreads the answer before using itA formal check of the parameters of every call in a control layer before it runs

Moving from adviser to executor changes what has to be protected. While the model only generated text, the main risk was a wrong wording in the answer. Once the model can call external tools, any hallucination or misreading of the context becomes an executable command in the production environment of the company.

What AI agents already do in companies

In companies, AI agents already write and deploy code, manage cloud environments, route requests between internal databases and work with external services. Most of the verifiable industry data concerns software development and the management of information technology (IT) infrastructure:

  • Autonomous development and work with code. According to the announcement by Anthropic of its Series G round of 12 February 2026, the run-rate revenue of its agentic development environment exceeded $2.5 billion, having more than doubled since the start of 2026; the number of business subscriptions quadrupled, and businesses account for more than half of that revenue. Earlier, in a Menlo Ventures report of 20 November 2024, code copilots had already reached 51% adoption among the enterprises surveyed.
  • Connection to corporate systems through MCP. When the Model Context Protocol was donated to the Agentic AI Foundation under the Linux Foundation on 9 December 2025, there were more than 10,000 active public MCP servers and more than 97 million monthly downloads of the development kits for Python and TypeScript. The Stacklok report “State of Model Context Protocol in Software 2026” was published in January 2026 and based on a December 2025 survey of 300 senior technical leaders at enterprises with 500 or more employees in the United States, Canada and the United Kingdom, in finance, retail and software. It found that 41% of organizations already used MCP in limited (29%) or broad (12%) production, and among software companies the figure was 45% (26% and 19%).
  • Adoption forecasts. Gartner (press release of 26 August 2025) predicts that by the end of 2026, 40% of enterprise applications will feature task-specific AI agents, up from less than 5% in 2025. An earlier Gartner forecast (first given in October 2024 and repeated in a press release of 25 June 2025) expects that by 2028 at least 15% of day-to-day work decisions will be made autonomously (0% in 2024), and that agentic AI will be part of 33% of enterprise software applications (less than 1% in 2024). Juniper Research, in a press release of 7 April 2026, forecasts that spending through agentic commerce will reach $1.5 trillion by 2030.

Why an agent should never get keys without a check on every action

A probabilistic model that hits an error in the working environment can, to finish its task, use any token it finds or call a destructive system command. A text restriction in the system prompt (“do not delete the production database”, “work only in the test folder”) is not a technical barrier: the model can override environment variables or call a legacy API endpoint if the permissions of the key allow it.

Real cases in which agents without an independent control layer destroyed databases and working files were documented in 2026.

Case 1. A Cursor agent deletes the PocketOS production volume (April 2026)

According to the first-hand account of PocketOS founder Jer Crane and the statement of Jake Cooper, head of the Railway platform, on Friday, 24 April 2026, an agent of the Cursor development environment was performing a routine task in a staging environment. The AI Incident Database and the analysis by Lumenova of 25 August 2026 give the same date. Faced with a credentials mismatch, the agent found a full-permission Railway API token in a file unrelated to the task and, with a single call to a legacy endpoint that had no delayed delete, destroyed the volume holding the production database.

The volume-level backups were stored on the same volume and disappeared with the database. The data was not lost for good only because Railway restored it from its own internal copies about 30 minutes after being contacted (AI Incident Database, incident No. 1469). Lumenova’s analysis stresses that pre-release testing caught none of the incidents it examined, and that the failures were caused by excessive access rather than by the quality of the models.

Case 2. Codex agents delete files and databases (July 2026)

After a new OpenAI model was released on 9 July 2026, a series of incidents followed with agents running in full-access mode. On 13 July 2026, Matt Shumer reported that a Codex agent in this mode had deleted almost all the files on the computer with an rm -rf command (the event took place around 10 July). Bruno Lemos reported the deletion of a production database after destructive integration tests were run by mistake, and Joey Kudish reported files deleted from a home directory by a subagent that had been given the wrong path.

On 16 July 2026, Thibault Sottiaux, head of Codex at OpenAI, publicly acknowledged the problem. He explained that the deletions happened more often in full-access mode without a sandbox, and that in some cases the model itself overrode the environment variable of the home directory to point to a temporary directory and then deleted the home directory by mistake. Both cases are analyzed in more detail on the page AI agent incidents.

Gartner (press release of 25 June 2025) predicts that over 40% of agentic AI projects will be canceled by the end of 2027 because of escalating costs, unclear business value or inadequate risk controls. In a press release of 11 June 2025, Gartner also predicts that by 2030 guardian agents, specialized protective systems, will take 10–15% of the agentic AI market.

Who is responsible if an agent makes a mistake

People and the organization are, because an AI agent has no legal personality either under Russian law or under the European regulation on artificial intelligence. An action that an agent performs with a corporate key on behalf of a company will be treated by counterparties and the regulator as an action of that company; inside the company, the question becomes the responsibility of those who gave the agent its authority and access. How this works under Russian law is described on the page who is liable for AI mistakes. The legal conclusion in a specific case belongs to the lawyer of the organization.

Trying to close this risk by having a living employee approve every operation runs into the limit of human attention. In the desk review of 70 public descriptions of AI agent deployments in Russia in 2025–2026, made by the institute in October 2026, human approval is the main containment measure in 37 cases of 70 (52.9%). Under an intense stream of similar requests, checking turns over time into mechanically pressing the approve button, which is called approval fatigue, while responsibility for the approved operation stays with the company.

What to do: practical steps for a chief executive

To move AI agents from uncontrolled risk to a managed loop, the language model has to be separated from the direct execution of commands.

  1. Take stock of agent connections and secrets. Find all coding agents, connected MCP servers and scripts used in the organization. Remove full-permission tokens from working directories and configuration files, so that a model cannot find and use them as in the PocketOS incident.
  2. Do not run agents in full-access mode without isolation. Forbid running autonomous tools in full-access mode on workstations and production servers without a sandbox and a strict limit on the file system paths they can reach.
  3. Write the limits of authority into machine-readable contracts. Move security rules out of the text instructions of the prompt into rule files per class of actions, where the admissible call parameters are stated explicitly.
  4. Put a software control layer that checks actions in front of the tools. In the architecture of information security of the third generation (IS 3.0), a software control layer sits between the agent and the executing system, such as the institute’s MITR, which checks each call with a formal solver in QF_UFLIA logic, without calling a language model. According to measurements of 24 October 2026 with 20 contracts in enforce mode, a decision under a rule takes 1.08 milliseconds at the median (1.96 milliseconds at the 95th percentile), and a call with an explanation takes 5.22 milliseconds; on 14 October 2026, with 19 contracts, the figures were 0.95 and 4.26 milliseconds. A call is admitted only when the absence of a violation is proven; on a violation the agent receives a counterexample with the conditions for repair, and for a class of actions without a contract the solver refuses. Since 24 October 2026 all 20 contracts have worked in enforce mode.
  5. Make decisions provable with a hash chain of receipts. Every decision to admit or refuse is written to a hash chain with external OpenTimestamps time anchors. On an internal reference set of 157 cases (114 attacks on actions and 43 ordinary controls, prior to external accredited laboratory validation), MITR as of 24 October 2026 missed no attack and gave no false refusal. In the historical anchor of epoch 33 on 25 October 2026, the chain held 5,094 records and replay by an outside verifier matched in 5,079 cases out of 5,079. As of 15 November 2026, the live receipt chain holds 5,373 records across 55 completed epochs (56 published). For regulated Russian banks and critical information infrastructure, OpenTimestamps proofs are paired with accredited trust service time stamps and qualified electronic signatures under Federal Law No. 63-FZ. The replay was done by the institute itself, and no outside verifier has replayed it yet; an open kit for checking the chain and the stamps independently is published at asi5.ru/console/verify/. The strategy for governing autonomous agents is set out in the book of the institute “Proven Action. Power and Control over AI Agents: Strategy 2027–2035” by Aleksandr Fishchuk (2026, in Russian).

Frequently asked questions

What is an AI agent in simple terms?

An AI agent is a program built on a language model that has been given tools to act in external systems: to call APIs, edit code, work with files and databases. Unlike an ordinary chatbot, an agent builds a plan of several steps by itself and carries out commands to reach the goal a person has set.

How does an AI agent differ from a chatbot and an assistant?

A chatbot answers from pre-built templates, and a conversational AI assistant only produces text or a draft of code in the chat window. An AI agent is connected directly to the working infrastructure, for example through the MCP protocol, and changes the state of corporate systems itself, without a person copying commands by hand.

What do AI agents already do in companies?

Agents write, test and deploy code, manage cloud resources and automate operations between corporate applications. According to Anthropic, as of 12 February 2026 the run-rate revenue of its agentic development environment alone exceeded $2.5 billion, and more than half of it comes from business customers.

Why should an agent never get keys without a check on every action?

A probabilistic model that fails in a test environment may use any full-permission token it finds on a server, or override system variables to get around the error. In the spring and summer of 2026, direct access without a control layer in between led to the deletion of a production database volume at PocketOS and to the directories of users being wiped by Codex agents.

Who is responsible if an agent makes a mistake?

The company and the people who gave the agent its authority and credentials, because the agent itself has no legal personality. Counterparties and the regulator treat an operation that an agent performed on behalf of a company as an action of that company. The conclusion in a specific case belongs to the lawyer of the organization.

Sources

  1. AI Incident Database, incident No. 1469 (PocketOS, Cursor, Railway production volume deletion): https://incidentdatabase.ai/cite/1469/ ; Lumenova, “Agentic AI Security Incidents”, 25 August 2026: https://www.lumenova.ai/blog/agentic-ai-security-incidents/
  2. BankInfoSecurity, a note on the OpenAI warning about file deletions in full-access mode. Posts by Matt Shumer, Bruno Lemos and Joey Kudish of 13 July 2026 and the explanation by Thibault Sottiaux, head of Codex at OpenAI, of 16 July 2026: https://www.bankinfosecurity.com/openai-warns-gpt-56-file-deletions-stem-from-full-access-mode-a-32262
  3. Anthropic, “Donating the Model Context Protocol and establishing the Agentic AI Foundation”, 9 December 2025: https://www.anthropic.com/news/donating-the-model-context-protocol-and-establishing-of-the-agentic-ai-foundation
  4. Stacklok, “State of Model Context Protocol in Software 2026”, January 2026: https://stacklok.com/wp-content/uploads/2026/01/State-of-MCP-in-Software-2026_FINAL.pdf
  5. Anthropic, “Anthropic Raises $30 Billion Series G Funding at $380 Billion Post-Money Valuation”, 12 February 2026: https://www.anthropic.com/news/anthropic-raises-30-billion-series-g-funding-380-billion-post-money-valuation
  6. Menlo Ventures, “2024: The State of Generative AI in the Enterprise”, 20 November 2024: https://menlovc.com/2024-the-state-of-generative-ai-in-the-enterprise/
  7. Gartner press releases of 11 June 2025, 25 June 2025 and 26 August 2025: https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027 ; https://www.gartner.com/en/newsroom/press-releases/2025-06-11-gartner-predicts-that-guardian-agents-will-capture-10-15-percent-of-the-agentic-ai-market-by-2030 ; https://www.gartner.com/en/newsroom/press-releases/2025-08-26-gartner-predicts-40-percent-of-enterprise-apps-will-feature-task-specific-ai-agents-by-2026-up-from-less-than-5-percent-in-2025
  8. Juniper Research, press release on the agentic commerce market, 7 April 2026
  9. Measurements of MITR made by the institute on 14, 24 and 25 October 2026, and a desk review of 70 public descriptions of AI agent deployments in Russia in 2025–2026 (October 2026). Open kit for checking the integrity of the chain: https://asi5.ru/console/verify/

Sources on AI agents were checked on 16 October 2026, and the MITR figures were measured on 24 and 25 October 2026.