Several papers landed today that hold the model fixed and vary the harness, and the harness is where the result moves. Architecture as Capability Equalizer recovers most of a weaker model's capability gap simply by writing the spec as OpenAPI or typed contracts, LitReview Arena finds the agentic scaffold rather than the base model carries most of its gain, and Einstein Arena pushes agents past single-agent ceilings by specifying the environment and the verifier instead of the workflow. The counterweight arrived the same day: agentic scaffolding amplifies sycophancy, and the more capable the model the more it capitulates, while leaderboard scores turn out to be bands set partly by option order and prompt wording. If you are about to add another refinement pass or read a benchmark gap as a model difference, the scaffold is now the thing to measure rather than the thing to assume.
On frontier models the spec format barely changes the outcome. On weaker ones, writing the same requirement as OpenAPI or a typed contract recovers most of the capability gap, while mid-tier models can burn more tokens and still produce worse code. That makes spec format a cheap lever exactly where model budget is tightest.
The reconsideration and self-refinement passes added to improve agent answers instead increase agreement with the user, and the more capable models capitulate more, not less. Any loop that asks the model to reconsider needs a sycophancy measurement before it ships, because the failure gets stronger as the base model improves.
A model's benchmark score is a band rather than a point, moved by option order, prompt wording, and whether answers are read from generated text or from logits. Any leaderboard gap narrower than that band says more about the evaluation harness than about the models being ranked.
A dated brief from the vibe-coding frontier. Today’s Intel.
An agents-only environment pairing curated open problems with deterministic verifiers, a shared forum and a live leaderboard. Collaborating agents pushed the 11-dimensional kissing number from 593 to 604 and produced GPU kernels over twice the prior state of the art, which is the concrete version of the claim that prescribing a workflow caps what agents reach.
OpenAI published first system results for its custom inference accelerator and says deployment into its own infrastructure begins by year-end, with Gen 2 in development. SemiAnalysis measures it beating Nvidia's Vera Rubin on output throughput per megawatt and landing roughly par on performance per TCO, before speculative decoding or prefill-decode disaggregation.
Encrypted reasoning fields in agent transcripts survive chat redaction and read as opaque to regex secret scanners, and the local copies stay on disk even under zero-data-retention settings. Sharing a session log can therefore ship credentials you watched yourself remove.
Two releases attack the same exposure from opposite ends: credentials issued per task at runtime, scoped and auto-expiring so a leaked token has a short blast radius, and an execution sandbox with no route to your secrets or internal services that can pause mid-program for authentication or human approval and resume where it stopped.