https://orbitalproof.com/blog/ Plain-English posts on crash recovery for AI agents and unsafe tool combinations, each tied to a published result. en Thu, 01 Oct 2026 00:00:00 GMT https://orbitalproof.com/blog/what-crash-recovery-means-for-ai-agents/ https://orbitalproof.com/blog/what-crash-recovery-means-for-ai-agents/ Thu, 01 Oct 2026 00:00:00 GMT Explainer An agent that dies halfway through a multi-step change can leave a mess, and a journal on disk is the usual way to make that change all or nothing. https://orbitalproof.com/blog/why-tools-safe-alone-are-unsafe-together/ https://orbitalproof.com/blog/why-tools-safe-alone-are-unsafe-together/ Thu, 01 Oct 2026 00:00:00 GMT Explainer Checking agent tools one at a time, or two at a time, can pass a combination of three that leaks a secret. https://orbitalproof.com/blog/why-hand-picked-crash-tests-are-not-enough/ https://orbitalproof.com/blog/why-hand-picked-crash-tests-are-not-enough/ Thu, 01 Oct 2026 00:00:00 GMT Explainer A recovery design that passes every crash point its authors chose can still fail at moments they did not think of. https://orbitalproof.com/blog/how-to-approve-agent-tools-as-a-set/ https://orbitalproof.com/blog/how-to-approve-agent-tools-as-a-set/ Thu, 01 Oct 2026 00:00:00 GMT Explainer Approving an agent's tools as a set means asking what the whole kit can do, and recording each approval so that it cannot be reused by accident.