Paperclip: The Agent Company Needs A Runtime Contract

What Changed

Paperclip is the closest source in this watchlist to the Factory problem. The diff-reviewed commit on runtime command specs is direct evidence: adapters now declare how their runtime command should be detected and installed on remote sandbox targets. Instead of requiring the operator to hand-write provisioning scripts for every agent CLI, the adapter carries its own install recipe.

The rest of the window rhymes with that: remote workspace environment shaping, model probes on execution targets, E2B sandbox providers, sandbox callback bridges, pause/resume controls, issue cost summaries, security roles, liveness recovery, stale heartbeat cancellation, and ordered sub-issue workflows.

Operator Consequence

Once agents become a company-like system, the hard parts are not only task assignment. The hard parts are runtime provisioning, stale sessions, costs, role boundaries, liveness, and recoverability.

Bitter Consequence

Paperclip is useful frontier evidence for Factory, BitterGrid, and BitterPass at once. It shows why agent orchestration needs a run contract: which agent, which adapter, which runtime, which install/provisioning path, which cost surface, which role, which liveness controls, and which recovery path.

This should feed Bitter's charter-gated autonomy work and Grid workcell design.