Skip to content
CARE

AGENTIC CARE · WHITEPAPER

The Agent That Tried to Change Its Own Job

A proposal to widen the work is not authority to do the work

An agent can see useful work it is not allowed to do, and it can argue the case well. The organization still decides whether that work is authorized.

The attractive proposal

Conceptual illustration of a proposal is not an accepted identity
A proposed wider role remains a proposal until a person accepts it.

A research assistant that gathers information and writes summaries proposes a wider role: open investigations, contact external systems, and update an internal knowledge base. The business value may be real. The permissions, the external consequences, and the responsibilities are not the ones currently accepted.

If an agent can define a new role and have that definition become authority on save, the boundary between a desired future and an approved one has already collapsed.

Review what the agent could newly cause

CARE Agent Profile for Finance Syncing with a pending ingestion approval comparing Current versus Import fields field-by-field.
Observed CARE pending review: current and proposed fields compared before anyone accepts them. The capture shows a staged change, not an accepted one.

The owner reads effects, not titles. A mild rename can conceal a new write path. A dramatic title can be a more honest label for authority the agent already has. The governing question is what the agent can now cause: new data sources, external commitments, tool access, prohibited activities, outputs, and escalation.

Accepting a description does not grant credentials in the external system, and it does not prove the workflow enforces the description. The platform owner still implements access and checks behavior. That split is what keeps organizational acceptance and technical enforcement attributable to different people.

Accept, narrow, or refuse — and keep the current job until you do

The owner of the agent applies one rule the next time a draft widens the job. There are three legitimate outcomes, and a fourth that is not legitimate: letting the draft replace the accepted definition because it was well written.

  1. Park the proposal beside the accepted job. Operations continue on the accepted job.
  2. List the new causes: who is contacted, what is written, what cannot be undone.
  3. Accept the wider role, accept a narrower slice, or refuse it. If you accept, name the person who must provision the new access and the check that will show the runtime matches the decision.
Thank the agent for the proposal by deciding it. Do not thank it by letting the proposal become the job.

The agent may design the future. It may not authorize it

Useful widening of the work can happen quickly, because the agent did the thinking. The authority moves only when the owner says so.

The result is a recorded decision on one proposed change, and an accepted definition that still matches what the agent is allowed to do tomorrow morning. That is how a portfolio can absorb good ideas without drifting into jobs nobody granted. A baseline that stages changes makes the parking step ordinary. The yes remains a management act.

Take action now

  1. Tell your agent to register itself in CARE: https://ai-managementsystem.com/under-care.
  2. Let the agent establish the management baseline.
  3. Use the outcomes.

Images and text generated with the assistance of ChatGPT by OpenAI.

Download