Prospero, a new coding harness, treats hangs and orphaned processes as the main failure modes

The author of the nullprogram blog released Prospero, a coding harness written in C++ with a Dear ImGui interface. It includes no model providers. Users supply their own: llama.cpp, an on-prem model, OpenRouter, or the Anthropic or OpenAI APIs. The post says a Claude subscription cannot be used. The author claims it uses about two orders of magnitude fewer resources than other harnesses; the post gives no measurements.

The stated design rules are the useful part. Under the heading "Never wait forever", every operation has a deadline, and a permission request nobody answers is denied by timeout so the agent carries on without it. A shell command that hits its timeout becomes a background command that the agent can keep waiting on, and the agent is woken at intervals to decide whether to continue. When a command is stopped, the whole process tree stops with it; the author says a regular annoyance earlier this year was a stopped build that kept running in part while the agent started a second one. Agents can spawn subagents and message each other, and a headless copy can be deployed to another host over SSH or WSL. The free Community Edition omits the built-in headless engines; a paid Professional Edition includes them.

The author also writes that the permissions system will not stop a malicious agent, and that containing one means running the harness itself in a container or on separate hardware.

Why it matters: If you run agents overnight, the post gives you three behaviors to test in your own harness: a deadline on every call, timed-out commands moved to the background instead of hanging, and cleanup of the full process tree. The efficiency claim is unmeasured and the project is a day old, so treat it as a design argument rather than a benchmark result.