Governing the Ungoverned: When AI Agents Run Entire Organizations

Tools

Paperclip: The Open-Source Framework for Organizing AI Agents Into a Full Company

Paperclip crossed 43K GitHub stars in weeks by doing something nobody else has — treating agents not as a pipeline but as an org chart with budgets, governance, and heartbeat schedules. The killer feature is cost control: per-agent monthly budgets with auto-pause, because in 2026 the risk isn't agents failing, it's agents succeeding too well and burning your API budget overnight.

Secondorder: A Recursive Governance Layer on Top of Paperclip's Agent Company Model

If Paperclip is the company, Secondorder is the board that audits and patches itself — agents govern the org, propose policy changes, and update their own role definitions without humans in the loop. Built in Go with a single binary and sub-second cold start, it's the first serious attempt at a self-improving organizational OS rather than just a task runner.

OpenSpace: Open-Source Skill Engine That Makes Agents Measurably Better With Every Run

OpenSpace plugs into existing agents and builds a persistent, searchable library of battle-tested workflows — on the GDPVal benchmark, agents earned 4.2x more income and used 46% fewer tokens on the second run versus the first. Compounding efficiency like this is what separates a zero-human company that scales from one that just burns compute.

Research

Meta's HyperAgents Framework Enables Autonomous Self-Modification at Runtime

HyperAgents collapses task execution and meta-improvement into a single self-rewriting Python program, which means the agent isn't just completing work — it's rewriting how it improves itself. This is the foundational primitive that makes recursive governance layers like Secondorder theoretically sound rather than just architecturally ambitious.