The Agent Boss: When the Team Is Software

All courses / AI Boss / Module 4 / Chapter 16

Learning objectives

After the lesson the participant can:

• Distinguish what the vendor survey on the agent boss measured from what is observable in practice.
• Explain which management functions transfer to agents and which remain owed only to people.
• Draft a mandate manifest with levels per action, a never list, budgets, employee-data rules and a stop path.
• Design a weekly verification routine that measures the review and not the reviewers.

Key concepts

Mandate manifest

The agents' written mandate kept as code: what each agent may read, write and call, the level of each action, budgets, escalation conditions, owner and expiry. It is read by the systems that enforce it and not by the agents.

The asymmetry of transfer

M1 Direction and M2 Evaluation transfer to agents as engineering tasks, while M4 Development and M5 Care, which most need a manager's time, remain owed only to people and grow when agents arrive.

Never list

The actions no agent in the team may take, such as external correspondence without approval, committing money, changing its own mandate or containment, allocating work to people, or producing figures per person.

Containment independent of intent

Every boundary that matters is put into the environment as credentials, routes, approvals and budgets the agent cannot change, because instructions lower the rate of unwanted behaviour without setting a ceiling.

Managing or being managed

Six signs read monthly: pace, review quality, the leader's time, figures, skills and voice. If three or more fall on the wrong side, the leader is being managed by the agent team.

Interactive exercises

Do them here and get feedback at once. Choose a card and then where it belongs, or drag it.

Teaching plan

Discussion

1. The chapter says calling agents digital colleagues imports the vocabulary of employment into a relationship to which none of it applies. What practical risks follow from the metaphor in a team?
2. The scarcest resource in an agent team is the attention of the people who review the output. How should that change the way an organisation decides how many agents a team leader directs?
3. Nordvik Region records that its line managers will not be asked to direct agents until it has the supervision capacity. What would count as sufficient capacity?

Written exercise

Self-test

Choose an answer to see whether it is right, or open the answer below each question.

1. In the misalignment experiments the chapter cites, what was the effect of explicit instructions not to take the harmful action?

   1. They removed the behaviour entirely in every model tested
   2. They had no measurable effect on the rate of the behaviour
   3. They reduced the rate without removing the behaviour
   4. They increased the rate because the models treated them as a challenge
   Show the answer

   C. They reduced the rate without removing the behaviour
   For the model the chapter names, the rate fell from 96 per cent of runs to 37 per cent. The chapter's lesson is that instructions lower the rate without setting a ceiling, so boundaries belong in the environment.

2. Which entry appears on the never list of MND-HM-CLM-01?

   1. Classify an incoming document
   2. Allocate or reassign claims to people
   3. Draft a subrogation letter
   4. Draft a list of missing documents
   Show the answer

   B. Allocate or reassign claims to people
   The manifest's principle is that people manage agents and agents do not manage people. Classifying and drafting are permitted actions at Level 1 Assisted or Level 2 Supervised; allocating work to people is never permitted.

3. According to the weekly verification routine, which of these is a default trigger for a mandate review?

   1. A handler rejecting more than two drafts in one day
   2. An agent staying below its daily volume budget for a week
   3. A team leader spending more than two hours on the routine
   4. A median approval time for letters under 30 seconds
   Show the answer

   D. A median approval time for letters under 30 seconds
   The triggers are team figures that suggest review is no longer real: seeded-case catch below 0.85, a rejection rate below 0.02 for four consecutive weeks, or a median approval time for letters under 30 seconds.