K
Kraus Labs
← Back to Teach

Paperclip: Managing Agents at Work

How Paperclip models, organizes, and runs teams of AI agents, and whether its ideas fit AI-assisted engineering

Mission

Mission: Paperclip

Status: assumed. Vinicius has not stated a mission for this study yet. This one was written from his profile and the study request. Confirm or rewrite it in the next session.

Why

Understand how Paperclip models, organizes, and runs agents (core concepts, architecture, and how you would actually use it), then judge whether its ideas fit AI-assisted software engineering work at Wellhub.

Success looks like

  • Explain Paperclip's core model (company, agent, issue, heartbeat, adapter, governance) from memory, and say how each maps to code in the repo
  • Trace one heartbeat end to end: trigger, adapter execute(), agent API calls, atomic checkout, run record
  • Run a local test instance and watch one agent pick up and finish one issue
  • Write a short, evidence-based fit verdict: which Paperclip ideas (atomic checkout, goal ancestry, budgets, approval gates, heartbeats) to adopt, adapt, or skip for Wellhub engineering workflows

Constraints

  • English lessons, grounded in primary sources only (repo README, docs, source, official site); cite what was read, never invent features
  • Already fluent in harness engineering, agents, and skills, so skip "what is an agent" basics
  • Short lessons that fit in one sitting

Out of scope

  • Building a custom adapter or plugin (revisit only if the fit verdict says yes)
  • Paperclip's cloud or multi-tenant deployment and business model
  • Rolling Paperclip out at Wellhub; this study ends in a judgment, not a migration

Lessons

Reference

Learning Records

  • Study Started Assumed Mission

    Study started on an assumed mission; agent basics already known

    Vinicius asked for a Paperclip study but gave no explicit mission, so [[MISSION.md]] is marked assumed: understand how Paperclip models and runs agents, then judge the fit for AI-assisted engineering at Wellhub. He already knows harness engineering, agents, and skills, so lessons start above the agent loop, at orchestration (who works on what, when, at what cost, with whose approval). They should not re-teach what an agent is.

    Evidence: stated prior knowledge in the study request (2026-09-26). No quiz results yet.

    Implications: Lesson 0001 teaches the control-plane mental model (company, agent, issue, heartbeat, adapter, governance). Next, in his zone of proximal development: trace one heartbeat through the source (adapter execute() → agent API calls → atomic checkout → run record), then a hands-on test-drive. Confirm the mission before lesson 0002.