← Articles
Safety

Skill governance for AI agents

As AI coding agents gain skills, plugins, and tools, a new question appears: which of them should a given task be allowed to use? Exposing every capability to every task is convenient and unsafe. Skill governance is the practice of granting an agent only the capabilities a task needs, for the duration of that task, and no more.

Avorelo Topic: Safety Topic: Capability Topic: Governance 2 min read

More skills, more surface

Every skill or tool an agent can reach is a capability that can be invoked, correctly or not. A documentation edit does not need shell access. A formatting pass does not need network calls. When all capabilities are available by default, the agent's effective reach on any task is the union of everything it can do, not the small set the task requires.

Govern by task, not by agent

The unit of governance should be the task, not the agent as a whole. The right capabilities for a task are determined by what the task is: its scope, its boundary, its risk. Granting those at task start and revoking them at task end keeps the exposed surface proportional to the work, and keeps the idle state at zero reach rather than standing broad access.

All skills, always
broad standing reach
Task-scoped skills
granted, then revoked

Governance is also provenance

Skill governance is not only about how many capabilities are exposed; it is about which ones are trusted. As agents pull in third-party skills, knowing where a skill came from and what it can reach becomes part of governance. A capability you cannot account for is one you should not grant by default.

Capability should be scoped to the task. Granted when the work needs it, revoked when the work ends, and accounted for by source.

How Avorelo helps

Avorelo tracks access scope per task. Capabilities are granted at task start and revoked at task end, so the idle state is zero access rather than standing broad reach. Each task sees the tools its work requires and nothing else, which keeps the exposed surface proportional and auditable.

Grant only what the task needs.

Avorelo scopes capability per task: granted at task start, revoked at task end. Idle state is zero access. Local-first.

Start free See how Avorelo works