Paperclip gives agents an organizational control plane: roles, reporting chains, tasks, approvals, budgets, adapters, and heartbeats. Installing it does not create a company. It gives you a place to coordinate external agent runtimes under human governance.
The setup I use as a mental model has three layers: Paperclip coordinates the work, a reviewed skill library shapes how agents perform it, and constrained experiments improve one measurable surface at a time.
Layer 1: Paperclip coordinates
Start with one company objective and two roles, not fifteen fictional employees. Give each role a narrow responsibility, a named supervisor, an approved runtime, and a budget.
Tasks should state the objective, inputs, allowed paths or systems, approval requirements, and evidence of completion. Agents wake for work through the configured lifecycle and stop between heartbeats; they are not continuous autonomous employees.
Use approval gates for hiring new agents, changing strategy, increasing budgets, external writes, and production releases. Budget controls are useful, but I do not promise perfect immediate cost containment or complete traces without testing the exact current implementation.
Layer 2: Skills define how work happens
I use gstack as an example of a configured workflow library, not a universal guarantee. Its skill inventory and command behavior change with releases. Verify the current first-party repository before naming or depending on any specific skill.
The durable pattern is to assign each stage a clear procedure:
- discover the repository and constraints;
- design before implementation;
- test behavior before changing it;
- review against the specification;
- verify the final candidate;
- release only after explicit approval.
No command name proves a destructive action is blocked or a deployment is safe. Repository rules, actual permissions, tests, and release evidence do that.
Layer 3: Experiments improve bounded surfaces
Autoresearch contributes the keep-or-discard loop: change one constrained input, run an objective evaluation, compare against a baseline, and record the result.
For a business system, I use it on internal prompts, retrieval settings, test fixtures, or templates—not unreviewed pricing, customer messages, payments, or employment decisions. The experiment needs a holdout, cost ceiling, reversibility, and a human owner.
A first operating cycle
- Define one company objective and success metric.
- Create an operator role and one specialist role.
- Give the operator a read-only discovery task.
- Review the resulting evidence and approve a small implementation task.
- Run tests and a separate review before integration.
- Capture one failure as a candidate skill improvement.
- Test that improvement against a frozen example set.
This is deliberately less exciting than an instant zero-employee company. It is also observable.
The real point
Paperclip can make ownership and work state visible. A skill library can make procedures repeatable. An experiment loop can improve measurable behavior. None of them removes the operator.
The human still chooses the goal, grants permissions, approves consequential actions, and decides when evidence is strong enough to ship. That is not a temporary limitation. It is the governance design.
