Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Coding

Paperclip: native Runner now default

Written November 10, 2026

Paperclip's November 6 release, v2026.1106.0, makes the native Paperclip Runner the default for new agents. Existing agents keep the runner they already have, and the old runner is still available under Advanced. The change matters most at the moment you create an agent. After that, it is easy to miss.

The problem

Paperclip is the control plane over the agents you already run: it organizes them, and it does not do the model work itself. Until now, a new agent started on a legacy adapter for its agent type: Claude, Codex, OpenCode, or Pi each had its own local adapter. The November 6 release changes that default, so a new agent starts on Paperclip's own native Runner instead. If you create agents the way you always have, you are now on a different runtime without having chosen it.

What changed

The release notes for v2026.1106.0 put it plainly: the Paperclip Runner is now the default for new agents. New Claude agents on Apple Silicon, Intel macOS, and Linux x64, new Codex agents on those plus Linux ARM64, and new OpenCode and Pi agents on macOS and Linux x64 now start on the native Runner instead of the legacy adapter. The switch applies across onboarding, the API and CLI, hires, built-ins, plugins, and imports. Linux needs GNU libc 2.39 or later. To keep the classic adapter for a new agent, open Advanced and set Runner to Legacy runner.

Existing agents are untouched: they keep their current runner. Existing Claude and Codex agents on qualified targets get a one-click Upgrade runner action that converts them, keeping the model, account, and instructions, and listing any settings that could not be converted.

Paperclip says Claude Runner agents start with their core task tools preloaded and were measured about 20% faster on simple tasks. That figure is Paperclip's own measurement from the release notes, not an independent test.

What else came in the release

The same release carries 515 commits. Alongside the Runner change, GitHub Copilot joins as a native agent option, running through the Paperclip Runner. Project repositories attached to a new project are now imported as skill sources automatically. A model-backed prompt injection check can classify tool results before they reach an agent, and agent logs and traces can be sent to Braintrust or any OTLP endpoint. Workspaces clean up after themselves: an hour after a task tree finishes, the reaper removes regenerable directories such as node_modules, dist, and target, keeping the branch and all source changes.

One breaking change to know about: the Pipelines experiment is removed. Pipeline boards, stage settings, case automation, review queues, the pipelines CLI commands, and the enablePipelines setting are gone; Cases and Routines remain, and existing pipeline tables and records are preserved untouched.

What to do

Limits

Related playbooks

Sources

Back to: Coding playbooks hub