A brief for policy
Every organisation shapes how its people behave, through pay, rules, roles and the habits of the place. Now part of the work is done by AI agents. Agentic Resource Management asks which of those levers still work, which fail and what has to be invented, in three settings: a person with an agent, an agent on its own, and agent to agent.
This page gathers what the site says for readers who shape training, certification, labour rules or public programmes. Each section links to the page it comes from.
Why the field exists
Labour rules assume the worker is a person who can be trained, certified and held responsible. Agents fit none of those assumptions neatly.
Example. A skills mission deciding what to certify has to ask which tasks a trainee will do, and which an agent will do under the trainee’s direction.
What the terms mean
The definition keeps accountability with a person or an organisation. It does not treat agents as legal persons.
Example. A rule on automated hiring can name the organisation that deployed the agent, rather than the agent, as the party that answers.
Who owns capability
People need consent and voice because no organisation owns their capability. Agents need neither, so accountability has to be assigned on purpose.
Example. Rules on consulting workers cover the people affected by a new system. They say nothing about who signs off the agent itself.
What could be measured and reported
Measures like these could become reporting standards, much as safety incidents are reported today.
Example. A public body could ask employers that deploy agents to report escalations, and errors that travel along chains of agents.
Agents on their own
Rules on automated decisions mostly assume one system making one decision. Agents that act over many steps test that assumption.
Example. An agent that collects documents, scores an applicant and sends a rejection has made three decisions. Which of them does a rule on automated decisions cover?
Chains of agents
Accountability rules written for one employer and one worker do not yet say who answers along a chain of agents run by different firms.
Example. If a bank’s agent relies on a credit agency’s agent, a regulator has to decide which firm answers for a wrong decision.
Where public institutions can act first
Standards, labour law and measurement are where public institutions can act first.
Example. A common format for reporting agent errors would let a state compare sectors without collecting personal data.
HR practices with no agent equivalent
The rows marked Open question are where existing labour rules have nothing to say about agents.
Example. Employee voice has no agent equivalent, so the question becomes how the people affected by an agent’s work are heard.
Accountability
Belief seven carries the most policy weight: someone must always answer for what a chain of agents does.
Example. A licensing rule could require every deployed agent to have a named, responsible owner inside the organisation.
Three processes with no agent equivalent yet
- Motivation and incentives. Reward shapes behaviour during training, not on the job; a deployed agent responds only to instructions, context and constraints. What plays the role of incentives is an open question.
- Norms and culture. Policies and written constitutions in an agent’s instructions; conventions that may emerge among agents that interact repeatedly.
- Employee relations and voice. Agents give no consent and have no voice, yet someone must answer for what they do; accountability has to be designed into the chain.
The beliefs, and the evidence so far
- ArguedManagement has always been about behaviour. Drawn from research on how organisations shape people’s behaviour. The comparison for agents starts with the From HRM to ARM table.
- ArguedAn agent is anyone, or anything, given a goal and discretion. A definition offered for the field to test and refine.
- ArguedIncentives move people. What moves agents? Drawn from how AI agents are trained. Experiments on incentive-like mechanisms are planned.
- PilotingA harness is an organisation. Controlled experiments comparing control layers in agent harnesses are underway.
- OpenNorms can be written down, and they can emerge. Simulations of agents interacting over many rounds are planned.
- PilotingPeople stay at the centre. ARTScore, which measures what people add to an AI’s best draft, has been tested with more than 150 professionals, and a randomised controlled trial is being set up.
- ArguedWhen agents manage agents, a person is still accountable. Drawn from how accountability works in organisations. Studies of delegation chains between agents are planned.
- ArguedTest it where work actually happens. A rule the field sets for itself rather than a finding.
Tested: checked against evidence. Piloting: being tried out now. Argued: reasoned from theory and existing research. Open: not yet known.
Questions still open
- Can incentives steer an agent that wants nothing, or only the people who train it?
- If an agent’s dispositions come from its training, who is its real employer: the lab that trained it or the organisation that deploys it?
- Can social norms be instilled in agents by instruction, and do they emerge among agents on their own?
- What does it mean to appraise a worker that can be copied, paused and rolled back?
- Is a model capital an organisation owns or a resource it rents, and what does that change about how it is managed?
- When an agent delegates to another agent, where does responsibility sit?
- What happens to human expertise when the routine work that once built it goes to agents?
- Which management practices transfer from people to agents, which fail, and what has to be invented?
Working with Venkat
Organisations can work with him on research collaborations, pilots and consulting. Policy partners are among the people he is looking for. Write to him, or email ramganuthula@gmail.com and ram@iitj.ac.in.