2 of 8

Three words, two kinds of worker

The thesis says management means shaping behaviour. Here is what that means, word by word, starting with the answer.

Working definition

Agentic Resource Management is the study and practice of directing, shaping and governing the behaviour of agentic actors, people and AI agents, that an organisation entrusts with goals and discretion, and of building, allocating and renewing the resources their work depends on, so that their combined work stays within intent and someone remains accountable for it.

Agentic

Agency is the capacity to pursue a goal by choosing and taking actions, and adjusting to what happens. It comes in degrees.

In people

Goals of their own
People pursue their own goals as well as the organisation’s. Principal–agent theory starts from this gap: interests, information and effort can diverge, and no contract specifies everything.
How they act
Under bounded rationality, and in response to incentives, norms, identity and relationships.

In AI agents

A model in a loop
An AI agent is a language model run in a loop by the software around it, the harness. It reads its context, chooses an action such as calling a tool, running code or messaging another agent, observes the result and repeats until the task is done or a stop condition is reached.
Goals given, dispositions trained
Its goals are given to it. Its dispositions come from pre-training and from fine-tuning and reinforcement learning on feedback; its behaviour at work depends on what is in its context and what the harness permits.
Autonomy by design
How autonomous it is depends on design: which tools it has, how many steps it may take alone, and where it must stop for a person.

In ARM, an actor is agentic when it is given a goal and discretion over how to reach it, so its behaviour has to be shaped rather than fully specified. That holds for an employee and for an AI agent; what differs is where the behaviour comes from and which levers reach it.

Resource

A resource is what work draws on, and what an organisation must acquire, allocate, protect and renew.

In people

Capability
Knowledge, skills, abilities and other characteristics, explicit and tacit.
How it forms
Through education, deliberate practice, experience on the job, socialisation and relationships. It sits inside the person.
Scarce inputs
Time, attention and judgement, including the time it takes to review an agent’s work.

In AI agents

Model weights
The trained parameters that hold an agent’s general capability. Costly to create, quickly superseded, and usually rented through an API rather than owned.
Compute and tokens
Inference is metered in tokens and processor time, with costs, latency and rate limits.
Context
The context window is the agent’s working memory for a task: finite, costly, and decisive for behaviour.
Data
Training data shapes capability; the organisation’s documents, reached through retrieval, give the agent its knowledge of the firm; logs and feedback drive evaluation.
Tools and permissions
APIs, code execution, browsing, files and credentials: what the agent can actually do, its authority.
Memory and the harness
Persistent memory, orchestration code, instructions, evaluations and guardrails, which the organisation builds and owns.

In ARM, the resource is not the worker. It is the capability an organisation puts to work, whether it comes from a person or from an agent, together with what that work consumes.

Management

Management is coordinating and shaping the behaviour of others towards shared goals when the work cannot be fully specified in advance.

In people

Levers
Selection, socialisation, incentives and pay, monitoring and feedback, rules and roles, culture and social norms, voice and consent.
When they act
Incentives act at the moment of work: people change what they do because of what they expect to gain or lose.

In AI agents

Levers
Model selection and evaluation on the organisation’s own tasks; fine-tuning; instructions, system prompts and written constitutions; what goes into the context; tool permissions and sandboxing; tracing, logging and evaluation; approval gates; versioning and rollback.
When they act
Reward shapes an agent during training, not while it works. At work, only instructions, context and constraints reach it.

In ARM, management means knowing which levers reach which kind of actor, and when each one works.

Resource or capital?

The name matters. “Human resources” has long been criticised for treating people as inputs to be used up. Human capital theory treats capability as an asset built through investment that earns returns over time; the resource-based view of the firm treats people as a source of advantage when their capability is valuable, rare and hard to imitate. ARM takes a position for both kinds of worker.

People: capital that belongs to the person

Human capability is built by investment in education, practice and experience; it lasts, grows with use, and is embodied, so it leaves when the person leaves. An organisation never owns it. It gains access through an employment relationship, which is why people need incentives, consent and voice. What the organisation manages is capability in use: time, attention and judgement given to a task.

AI agents: capital that can be owned, copied or rented, running on resources that are used up

Model weights, harnesses and the organisation’s data are capital: costly to build, but copyable at almost no cost and quickly superseded. Compute, tokens, context and human review time are flow resources, consumed task by task. An organisation can own or rent agent capital and needs no consent to deploy it, but it does not fully control the behaviour that capital produces.

Operational definition

ARM is in play whenever an organisation:

  1. delegates a goal, with discretion over how to reach it, to a person, an AI agent or a chain of agents;
  2. allocates resources to that work: human time, attention and judgement; compute, tokens, context, data and tool access;
  3. shapes behaviour and checks outcomes through levers such as selection, instructions, incentives or reward, permissions, monitoring, evaluation and approval.

The unit of analysis is the delegation: who gave which goal to whom, with what resources and discretion, and who answers for the result.

What can be measured:

  • Intent fidelity: how closely the outcome matches what was intended
  • Tokens and time to outcome: what the work consumed
  • Human review load: escalations and the time spent approving
  • Error propagation: how mistakes travel along chains of agents
  • Elevation: what people add beyond the best the AI can do

Next, Dimensions maps the territory where this definition applies. Dimensions