3 of 8

Managing behaviour in three settings, and the ground they share

The operational definition said ARM is in play whenever a goal is delegated with discretion. This page maps where that happens: three settings, and the ground they share. Each dimension comes with the question that drives it.

A person with an agent

How people shape the behaviour of the agents they work with, and how working with agents changes people.

Delegation and intent

How people turn goals into instructions and context an agent can act on, and where its discretion begins.

How much meaning is lost between what a person wants and what an agent does?

Evaluation and elevation

Appraising an agent’s behaviour and output, and measuring what people add to it.

What does good performance mean for an agent, and for its manager?

Human capability

The skills people need to direct agents well, and the curiosity and judgement worth protecting.

Which skills rise in value, and which quietly erode?

Interfaces and experience

How people and agents see the same work, the same pages, the same data.

What should an agent-facing interface look like?

An agent on its own

What drives an agent when no person is watching each step, and the limits its harness sets.

Dispositions and training

Where an agent’s tendencies come from: pre-training, fine-tuning and reinforcement learning on feedback, all before it starts work.

Which of an agent’s behaviours were set by its trainer, and which can the organisation that deploys it still change?

Autonomy and scope

The tools an agent may use, how many steps it may take alone, and where it must stop for a person.

How much discretion should an agent have, and for which tasks?

Drift and specification gaming

An agent pursuing a proxy, such as a metric or a test, instead of the intent behind it, or drifting from that intent over a long task.

How do we notice when an agent meets the letter of a goal but not its purpose?

Oversight and control

The levers that bound an agent’s behaviour: checkpoints, permissions, sandboxing, tracing and the ability to stop.

Where should a person step in, and how often?

Agent to agent

How agents coordinate, delegate and check one another, and which norms and accountability hold without a person at every step.

Orchestration

Roles, hand-offs and structures when many agents work on one goal: organisational design written into the harness.

Which team shapes work for agents, and which fail?

Accountability chains

Tracing every agent action back to a responsible person.

Who answers when the fourth agent in a chain gets it wrong?

Trust and verification

Agents checking agents, the conventions that form among them, and how errors spread through a system.

When can one agent rely on another’s work?

Lifecycle

Provisioning, versioning, memory and retirement of agents.

What does an agent’s career look like, from first task to last?

Across both

The levers and conditions that apply to people and agents alike, in every setting.

Incentives and motivation

Incentives act on people at work; reward shapes an agent only during training.

What plays the role of incentives for an agent at work?

Norms and culture

Norms written into an agent’s instructions, and conventions that may emerge among agents that interact repeatedly.

Can norms be instilled by instruction, and do they emerge on their own?

Resource economics

Compute, tokens, context and human review time: the resources agent work consumes, in every setting.

How do we get the most outcome for the least spend?

Organisation design

Org charts, roles, reporting lines and harnesses designed for people and agents together.

What does a team look like when half of it is software?

Policy, ethics and labour

Law, rights and institutions for a mixed workforce.

Which rules keep people protected and accountable?

Measurement and standards

Shared ways to measure agent behaviour and outcomes, and human contribution.

What should every organisation be able to report?