Overview
Paperclip is a control plane for organizing AI agents into companies whose goals, roles, tasks, runs, costs, approvals, and evidence remain connected. The series moves from the company model and adapter-backed heartbeat to atomic ownership, durable wakeups, bounded recovery, budget enforcement, human decisions, audit, deployment posture, low-trust containment, and governed tools. Its value is governable continuity: autonomous work can remain attributable and inspectable across heterogeneous providers and interruptions. The boundary is equally important: Paperclip governs the paths it admits and observes, but cannot guarantee provider correctness, strategic wisdom, complete external recovery, or host-wide control.
Learning outcomes
- Paperclip coordinates goals, roles, tasks, runs, costs, and controls without owning each agent’s internal reasoning.
- Company-scoped ownership and goal ancestry keep agent work connected to a reason and a responsible next step.
- Finite heartbeats, durable wakeups, and bounded recovery preserve continuity without guaranteeing provider correctness or progress.
- Budgets, approvals, audit, deployment, and tool policies constrain managed effects, not unmanaged external processes.
Keywords
Preview every chapter · Unlock full text, audio, and video from chapter 4
PlansFAQ
What is Paperclip used for?
Paperclip is used to coordinate and govern work performed by multiple AI agents inside a company-scoped control plane. It links organizational purpose to execution, evidence, cost, and human intervention while leaving provider internals outside its core.
How does Paperclip connect an AI task to execution?
Paperclip connects a task to execution through company and goal context, an assigned agent, a durable wakeup, a finite heartbeat run, and an adapter that invokes the provider. The run records what the adapter reports without equating a provider result with business success.
How does Paperclip limit autonomous work?
Paperclip limits managed work with atomic ownership, budgets, pause controls, approvals, deployment rules, trust policies, workspace and secret boundaries, and call-time tool decisions. These controls apply to the execution paths Paperclip admits and supervises.
Can Paperclip guarantee that an agent’s work is correct?
No. Paperclip can preserve attribution, evidence, cost, decisions, and failure paths, but it cannot prove that a provider’s reasoning was sound, a remote side effect was desirable, or an external artifact was fully restored.