Today's GitHub Trending tells a clear story: agents are moving from single-shot prompts to multi-phase, long-running workflows.

Let's look at two repos:

cloudflare/security-audit-skill : A coding-agent skill for multi-phase security audits with independently verified, machine-readable findings. The key word here is multi-phase — the audit is broken into stages, each producing verifiable output.

NVIDIA/OpenShell : A safe, private runtime for autonomous AI agents, built in Rust. It isolates agent execution to prevent runaway operations.

Both repos address the same underlying problem: as agents take on longer, multi-step tasks, the risk of failure compounds . A multi-step workflow where each step has a high success rate still has a significant chance of failing somewhere along the chain.

The Checkpoint Problem

Here's a scenario: imagine a security team runs an agent-based audit.…