The future of work with AI agents

At AR Labs we think AI will make people rethink the nature of work. In the picture we work toward, AI agents (language models run in a loop with tools, so they can act, not only reply) take on menial tasks and, over time, defined organisational roles, while people supervise them: setting direction, framing the work, judging results and deciding what should be built at all. We see three things that picture would ask for: interfaces built for supervising agents rather than for typing commands, organisations whose roles and decisions are written down precisely enough for an agent to fill them, and people who keep their own learning and judgement in working order.

Our outlook on where work is going

We hold a long-range hope, not a forecast: that menial tasks could largely be done by technology, through interfaces that are pleasant to use; that selling hours of one's time for money could stop being the basic shape of work; and that new sustainable technologies could ease much of the scarcity that work has been organised around. We hold this outlook because of what we see in our own practice. Once a task is specified well enough, an agent can carry it, and our working thesis, set out below, is that much of an organisation's work can be specified that way.

We do not know whether, how or when such a change would come, and we take the cost of getting there seriously. The essay series on our blog asks who will become an experienced engineer if AI replaces junior roles, and argues that some work, learning above all, should not be accelerated.

A company as software

Our working thesis for organisations is that a corporation is a kind of software: a set of roles, information flows and decisions that can be specified precisely enough for AI agents to run important operations. We test the thesis on ourselves. We run as a one-person company whose departments, such as engineering, content and research, are organised so that agents can take them over one at a time, with a strategy agent that reads status from each department and writes direction back. The agentic organisation describes that design.

Our plans go further. They describe a continuous model of the organisation's state, kept current by agents, replacing much of the coordination that a management hierarchy does by hand; capabilities that agents can select and combine for themselves; and agents that respond to events as they happen instead of waiting for a weekly review.

flowchart TD
  H["Person: direction, framing, judgement"] --> S["Strategy agent: reads status, writes direction"]
  S --> D["Department agents and coding agents"]
  D -->|"status up"| S
  D -->|"needs a person"| H
  S -->|"decisions beyond its authority"| H

What it asks of interfaces

If agents do more of the work, more of the person's job becomes supervision. In our own work, a person running several agents still finds and steers them by switching between terminal windows. Our answer is an interface built around the work itself. In VOX OS, our design for an operating layer over agent work, the dependency tree of all the work is the main surface, voice is the main input, and each piece of work can be opened as a live agent context that a person steers, forks or closes. What waits on a person is gathered in one place, so the person attends to decisions rather than hunting for them.

We draw a line through this stack. We expect much general agent tooling to become commodity, and put our effort into what we expect to stay hard: governance, a person's view and control of agent work, and interfaces that fit how people think. Here that means building the presentation and orchestration layers through which a person sees and steers the agents' work, and renting the rest: agent sessions, the task database, networking and code hosting come from existing tools. VOX Torch and Vox Canvas are the first surfaces of this design.

What it asks of organisations

An agent that starts each session fresh cannot rely on memory and meetings. So the layer between strategy and tasks, where approach is chosen and decisions are recorded with their reasons, has to exist as documents and records an agent can read. We think the same discipline helps the people in an organisation too.

Two further requirements follow from our experience:

• Ordering is the core of orchestration. Every actor, human or agent, needs a list of work in the right order; the page the intelligence is in the ordering develops this.
• Authority has to be explicit. Agents need written limits on what they may decide, a rule for when to act or ask, and checks that stop them when they exceed those limits. We extend trust one agent at a time, and a stage must succeed before the next begins. Our aim is a system that does the right thing when nobody is watching, and can show its work.

What it asks of people

We think people working with agents need more judgement, not less. The essay series on our blog urges people to slow down, think about what they are building and why, learn to say no, and keep a limit on how much of their attention and their lives they hand to AI. It warns that if AI replaces junior roles, the path by which people become experienced disappears, and that some work, especially learning, should not be accelerated at all.

For our own practice this means learning by building, with the rule that AI should not outrun the builder's understanding; the page the AI interface architect describes the kind of practitioner we think this calls for.

What exists today

We run our own work this way at small scale. A strategy agent runs planning reviews from daily to quarterly through shared status and direction documents, more than one long-running agent works independently, agents hand work to each other through a shared mailbox, and coding agents carry tasks from a dependency graph through to reviewed code that a person merges. Running several coding agents at once, agents that survive the laptop being closed, and voice control of the whole system are planned, not built.