Prospero puts deadlines and process cleanup into a coding harness
Chris Wellons released Prospero, a C++ coding-agent harness with a native interface, operation deadlines and explicit process-tree cleanup. Long shell commands become background tasks rather than being partly killed when a wait expires. It supports local or hosted model endpoints, but its author stresses that harness permissions are reminders and mistake checks, not isolation from a malicious agent.
Chris Wellons · nullprogram · Oct 31, 2026
Why it matters
Agent reliability depends on what happens when commands stall, processes outlive their task or permission requests go unanswered.
More context
The October 31 announcement offers a free Community build and a paid Professional build with bundled remote engines. Supply your own model endpoint and credentials. Try a disposable local project and inspect timeout behavior before giving it sensitive files or remote access.
Keep in mind
This is a new project whose performance claims were not reproduced here. Model usage is separate, and restricting an agent requires restricting the harness itself; the Professional price was not verified.
Original read