A conversation on the subject of this page starts with a message to the scientific director of the institute.
Write to the scientific directorThe future with agents: what changes in work, money and law, and what to do now
Programs built on neural networks no longer only answer questions: they make payments, write and ship code, carry on correspondence and change data. What follows from this for work, money, law and the state depends on which of three branches development takes to 2035; which branch is coming true is checked not by opinion but by external benchmarks with dates. This page is written for a person who is worried about it: without terms where they can be avoided, and with named sources wherever a figure stands.
A forecast here is not a promise: each names a branch and a benchmark by which it can be refuted. The rule is set out on the page Truth marking.
Three branches to 2035
The institute’s book Proven Action lays out the horizon to 2035 in three branches and takes the second as the working baseline; the conclusions for management are drawn so as to hold in any of the three.
| Branch | What happens | What an ordinary day looks like |
|---|---|---|
| A, “normal technology” | Model capabilities grow more slowly; artificial intelligence spreads over decades, as electricity and the internet did, mostly complementing human labor. | An agent takes over part of the routine, but decisions and most actions stay with people. |
| B, the trend continues (baseline) | Training compute grows four to five times a year, the price of a model call at the same quality falls roughly tenfold a year, and the horizon of tasks an agent completes without a person grows from hours to weeks. | Execution, such as correspondence, documents, payments and code, moves to agents; a person sets the tasks, signs the rules and answers for the result. |
| C, acceleration | Systems automate the cycle of their own research, and a decade holds a century of progress. | The density of machine actions per person grows by orders of magnitude; without a check of every action, control is lost. |
Which branch is coming true is visible from the benchmarks kept in the register of forecasts (fifteen indicators with sources, in Russian). The price of a model call at a fixed quality falls roughly tenfold a year (a16z, 12 November 2024; Epoch AI, 12 March 2025), and the task horizon of the best model is about twelve hours (Model Evaluation and Threat Research (METR), 20 February 2026). The expected arrival of an automated coder is given by the medians of two authors of the AI Futures Project: November 2027 and January 2030 (update of 16 August 2026). The electricity consumption of data centers is forecast by the International Energy Agency at 945 terawatt-hours by 2030, and the share of work decisions taken by agents is forecast by Gartner at 15 percent by 2028 against 0 percent in 2024. The details are on the page Horizon 2027–2035.
What changes
Work
Agents take over operational execution, and hiring shows it first: employment of specialists aged 22 to 25 in occupations exposed to artificial intelligence fell by 19 percent (Stanford Digital Economy Lab, revision of 12 August 2026; minus 13 percent in the first edition of August 2025). What remains is what a program is not given: setting the task, signing the rule, judging the result and bearing responsibility for it. In any of the three branches value shifts from the ability to execute to the ability to formulate, check and answer.
Money
An agent pays under a mandate: a person sets in advance what may be bought, for how much and on what conditions, and the seller and the payment chain receive a signed proof of those powers. The first live payment in Europe executed by an agent was carried out by Santander and Mastercard on 2 March 2026; the Agent Payments Protocol (AP2) and x402 protocols are described on the page The agent economy. For an ordinary person or a company this means: the agent needs a separate payment instrument with a limit, rules set before execution, and a trace of every payment in a log that cannot be corrected after the fact.
Law and the state
An agent is not a party to civil relations: the organization or the person on whose behalf it acts answers for its action. A director answers to the organization under Article 53.1 of the Civil Code of the Russian Federation (Who is liable, Director liability). Russia’s federal law on artificial intelligence was signed on 26 July 2026, and part of its provisions take effect on 1 March 2027; the draft amendments to Order No. 117 introduce measures for models and agents in state systems from the same date (section Rules and liability). The common direction of the rules is one: a program’s action must leave a trace that a reviewer checks without trusting the supplier.
What to do now
| For whom | What to do within a year |
|---|---|
| An owner or a chief executive | Decide which actions may be entrusted to agents and which only with a check before execution. Keep the customer base, memory, keys and the rules of transactions in a perimeter of one’s own, and keep a log of decisions that cannot be rewritten. How this is done: Formal verification; where to start: Custom automation. |
| A specialist | Learn to set a task for an agent and to check the result. Understand what an agent is and how it differs from a chatbot (What an agent is, the eleven-step learning path), and move one’s value into the part of the work where a rule, judgment and responsibility are needed. |
| An official or a regulator | Require from systems with agents not assurances but a verifiable trace of actions, and check requirements against the rules that are already taking effect. See how provability of an action is implemented in code: The doctrine of a managed transition and The program. |
What this page does not promise
- That branch B in particular will come true. It is taken as the baseline, and the benchmarks are checked once a year; branch A remains possible, and the conclusions for management hold in it too.
- Effective dates of rules beyond those named in the acts themselves: the dates were checked against primary sources on 25 October 2026 and are revised when the acts change.
- That action control protects against everything. A program’s action is checked against a rule that a person signed; the quality of the rule remains with the person.
Where to go next
For twenty minutes: Horizon 2027–2035 and The sixth technological order. For hours: The doctrine of a managed transition, The program and the book Proven Action: a map of 2027–2035 along five axes, three scenarios with benchmarks and a strategy for a person and an institute.
Autonomous Non-Profit Organization (ANO) “Research Institute of System Synthesis”, edition of 26 October 2026. Forecasts are marked with a branch and a benchmark; figures carry a date and a source.