B1: AI Futures 2035: Agents, Abundance, and Control

Reading

An AI agent swarm is a team of AI systems working on one goal. One agent may search. One may write. One may test. One may check safety. This matters because AI is moving from "a tool that answers" toward "a system that acts."

But swarms are not the main story. The bigger question is this: What happens when thinking, planning, coding, researching, designing, and checking become much cheaper? What happens when a student, a teacher, a clinic, or a small business can direct a small digital team?

The future is not one clean prediction. It is a fight between visions: abundance, race, human-centered design, and control.

So the serious question is not "How many agents can we run?" It is "What future are these agents helping us build?"

The serious AI future did not begin with chatbots. It began with an older dream: machines could help humans think better.

In 1945, Vannevar Bush imagined a "memex," a personal knowledge machine for connecting information. In 1960, J. C. R. Licklider wrote about "man-computer symbiosis." Symbiosis means two sides working together. He imagined humans and computers making decisions together: humans set goals and judge results; computers prepare choices and handle routine work.

In 1962, Douglas Engelbart wrote about "augmenting human intellect." To augment means to make stronger. In 1986, Marvin Minsky described intelligence as a "society of mind": many smaller parts working together. That idea helps us understand agent swarms today.

These older visions give us a test for 2035. If AI only creates more cheap content, it is not real progress. If AI helps people learn faster, find evidence, and solve harder problems, then it follows the augmentation dream.

Agent swarms matter because they change AI from conversation to action. A chatbot answers one request. An agentic system can break a goal into tasks, use tools, and check progress.

Current systems already show this pattern. Anthropic describes a multi-agent research system with an orchestrator-worker design: a lead agent creates subagents, sends them to search in parallel, and combines the results. Anthropic also warns that this can cost much more, so it is best for wide, valuable tasks.

OpenAI's Agents SDK uses handoffs. A handoff means one agent passes work to another specialist. Microsoft has shown Magentic-One, a research system with an orchestrator and specialist agents for web search, file reading, coding, and terminal work.

This does not mean "AI will do everything." It means many people may work by directing systems. They will set goals, check evidence, and decide when results are good enough. The future worker may be more like an editor, investigator, manager, designer, and judge.

Bill Gates argues that agents may become a new interface for computers. Instead of opening many apps, a person could tell an agent what they want in normal language. The agent may know the person's schedule, work, habits, and preferences.

The beautiful side is access: a student could get a tutor, and a small business could get planning support. The dangerous side is privacy. A personal agent is powerful because it knows you.

Sam Altman imagines a "gentle singularity": AI may make science and productivity much faster while daily life still includes family, nature, friendship, and meaning. Dario Amodei imagines AI speeding up biology and medicine. He calls one optimistic idea a "compressed 21st century," where decades of progress could happen in five to ten years.

But another vision is a race. Leopold Aschenbrenner argues that advanced AI may become an industrial and national race, driven by computer clusters, chips, electricity, security, and government interest. Whether his timeline is right or wrong, one idea matters: AI is becoming infrastructure.

Work data points in this direction. Microsoft's 2025 Work Trend Index says many workers may become "agent bosses." It reported that 28% of managers were considering AI workforce manager roles, and 32% planned AI agent specialist roles. The World Economic Forum says this decade may create 170 million jobs, displace 92 million, and change 39% of key job skills by 2030.

Not every serious thinker wants a machine-first future. Fei-Fei Li and Stanford's Human-Centered AI Institute argue that AI should improve human life and human capability. The key question is: "Does the system make people more capable, more informed, and more free?"

This vision asks practical questions. Does AI help teachers teach? Does it help doctors listen? Does it give workers more power, or only watch them more closely?

Technology writer Kevin Kelly gives a useful word: protopia. It means a future that becomes a little better than today while still creating new problems.

That is why protopia is a useful test for AI. The question is not only whether a system feels impressive. The question is whether it leaves people with more knowledge, more agency, and better judgment after the machine is gone.

A human-centered future asks people to become better question askers, evidence checkers, creators, and citizens, not only faster users of tools.

Mustafa Suleyman, co-founder of DeepMind and CEO of Microsoft AI, uses the phrase "the containment problem." The question is: How can society keep powerful technologies under control while still using their benefits?

This is not a small technical detail. If agents can write code, move money, send messages, or control machines, then permissions, monitoring, security, and human approval matter.

Google DeepMind's AI Control Roadmap uses a strong image: a driving instructor with dual controls. For agents, this means limited access, sandboxes, behavior checks, and supervisors that can stop risky actions.

The International AI Safety Report 2026 gives another warning. It says AI is improving quickly in math, coding, and autonomous operation, but performance is still "jagged." A system may be strong at a hard task and weak at a simple one. The report also notes at least 700 million weekly users of leading AI systems.

So control is not the opposite of the future. It is part of the future. Strong agents need stronger brakes. Personal AI needs privacy rules. The more power a system has, the clearer human responsibility must be.

By 2035, we may live inside several futures at the same time.

The first is the abundance future. AI helps scientists, teachers, doctors, artists, programmers, and small businesses do more.

The second is the coordinator future. Many workers manage agents. They set goals, review evidence, edit results, design workflows, and take responsibility.

The third is the control future. Companies and governments fight over data, chips, energy, models, and rules. Trust becomes a public issue.

The fourth is the protopia future. Communities use AI to make daily life a little better: a school improves lessons, a town translates services, and a student builds something real.

Technology will not choose between these futures by itself. People will choose.