Agents and delegated authority
An autonomous agent is software that can observe context, choose steps and pursue a goal with some degree of independence. Autonomy describes how a system operates; it does not settle what the system is permitted to do.
Define the agent in context
Agent behaviour can range from drafting a suggestion to calling tools or interacting with external services. A useful description names the model, tools, identity, operating environment and human checkpoints relevant to the task, rather than using “agent” as a guarantee of independence.
- What can the agent read or change?
- Which tools and credentials can it reach?
- Can it act repeatedly without review?
- What records let an operator reconstruct its decisions?
Separate capability from permission
A tool being available to software does not mean every use is authorized. Policies should constrain the action at the system boundary, not rely only on the model to remember a written instruction.
Good oversight includes narrow permissions, monitoring, an interrupt path and a plan for partial failure. People should know when they are approving a proposal and when an external action has actually completed.
Related foundations
Explore identity, authorization and mandates as separate questions in an agent system.