Source note
A case for an Autonomy Kernel
Summary
The paper argues for an autonomy kernel: a stable runtime under autonomous agents that controls authority, records actions, preserves state, and lets a principal stop agents at any time.
Problem
- Current agents are often prompt-bound sessions, yet users increasingly ask them to act for days with real authority.
- Agent tools lack a shared layer for naming running work, granting scoped permissions, auditing actions, preserving continuity, and revoking power.
- This matters because long-running agents need accountability, stoppability, and portability across models and vendors.
Approach
- The core model has 3 layers: the agent is the process, the model is the reasoning engine, and the kernel is the runtime underneath both.
- Authority starts with 1 principal. Purpose flows through principal, intent, goal, task, process, and action; power flows through policy, capability, lease, and syscall.
- Before any action runs, the kernel checks whether the action has both a traceable purpose and an explicit grant of authority.
- The kernel stays small and owns 5 mechanisms: execution, identity, authority, communication, and auditing.
- Memory is a curated, erasable projection over the audit record; the audit record is the durable source of truth.
Results
- The excerpt reports no benchmark, dataset, implementation result, or quantitative comparison.
- It proposes 9 design claims for autonomous systems, including disposable agents, replaceable models, traceable intent, explicit authority, full auditability, principal sovereignty, governed memory, and a small kernel.
- It defines 3 standardization commitments: a stable versioned boundary, portable state with guaranteed exit, and non-transferable accountability.
- It claims every action must map to 1 principal and 1 recoverable authorization path before execution.
- It claims adoption can be additive: existing tools should run above the kernel without a rewrite, but the excerpt gives no adoption data.