Notes RSS feed

The Wire — short items from the day's overflow, verdict follow-ups on earlier calls, and changelogs for the paper's maintained data assets. When a story earns a full argument, it runs as an article instead.

Wire

AWS Auth skill flags two destructive updates
08/08/26, 09:13 PM

AWS added one aws-auth skill to its Agent Toolkit, covering Cognito setup from OAuth 2.0 and passkeys through JWT authorizers, while the published skill carries two full-replacement warnings: update-user-pool-client and set-identity-pool-roles can erase omitted settings. The AWS launch notice says agents can pair the skill with IAM guardrails and CloudTrail logging through AWS MCP. For builders following the archive’s case for putting policy inside the agent harness, this is a useful reminder that procedural knowledge must encode destructive API semantics, not merely generate valid commands.
Wire

AgentOPSD reaches 89.1% without extra rollouts
08/08/26, 08:21 PM

AgentOPSD reached 89.1% success on ALFWorld with Qwen2.5-7B while adding neither a critic nor extra rollouts. The new preprint converts teacher-student log-probability gaps into turn-level evidence, recursively updates a Bayesian belief across an episode, and reports gains over GRPO and self-distillation baselines on ALFWorld, WebShop, and Search-QA; its linked repository says the full training code is still forthcoming. For teams extending the evidence that harness design changes agent outcomes, the method is worth filing as a potentially cheaper way to credit pivotal actions, but the score should remain provisional until the code enables reproduction.
Wire

OpenAI absorbs NextSlide after roughly one year
08/08/26, 08:20 PM

OpenAI acquired NextSlide after the presentation startup had operated for a little over 1 year, moving its team onto ChatGPT. Founder Ahmed Beshry’s announcement says NextSlide turned prompts, notes, documents, or research into editable presentations; TechCrunch reports that the deal actually closed earlier this year and its financial terms remain undisclosed. For builders tracking OpenAI’s shift from a model vendor toward an agent platform, the quiet acquihire is another signal that artifact creation—this time slides—is becoming native assistant capability rather than a separate destination app.
Wire

SkillZip compresses agent skills 3.46x
08/08/26, 07:27 PM

SkillZip compressed agent skill libraries 3.46x while preserving 99.2% of dependencies and 98.7% of verifier reachability, according to the framework’s new preprint. The authors report gains of up to 12.2 points over the strongest baseline and stable retrieval as libraries scale from 200 to 100,000 skills by replacing repeated procedural motifs with reversible, contract-bearing macros. For builders measuring the cost swing created by an agent harness, this suggests compressing executable structure rather than flattening procedures into shorter prose.
Wire

Activity Frames compress agent memory 86x
08/08/26, 07:26 PM

Activity Frames compressed a day of screen activity 86x in 68 milliseconds, with an agent answering questions from the result at 98.4% accuracy, according to a new preprint and open evaluation harness. The deterministic, zero-model compiler processed 128,756 frames across 51 active days, though all came from one professional, so the striking result is a single-user proof rather than a population benchmark. For teams following Cognition’s bet on persistent agent memory, the useful design clue is to preserve typed events and evidence pointers before paying a frontier model to summarize them.
Wire

Pre-commit gates lift self-evolving agents to 72%
08/08/26, 07:25 PM

Verifier-as-Gatekeeper reached 72% pass@1 on Terminal-Bench 2 with a skill pool roughly 5x smaller than unconditional accumulation, according to the authors’ new preprint. Its three critics screen each learned skill for structural validity, behavioral harmlessness, and semantic consistency before admission; the study also found that removing a defective skill later recovered only a small part of the damage inherited by descendant skills. For operators building the kind of runtime policy layer described in HiddenLayer’s agent-harness architecture, the fact to file away is that learned skills need a pre-commit gate, not merely a rollback button.
Wire

Amazon's GW Ranch carries a 33.2M-ton ceiling
08/08/26, 06:16 PM

Amazon confirmed that its planned Pecos County AI campus will buy power from GW Ranch, whose permit application calculates a maximum potential of 33.2 million tons of CO2-equivalent a year across 35 gas turbines. Cleanview’s investigation linking Amazon to the site stresses that plants rarely reach their permitted ceiling; the project developer describes 7.65 GW of gross gas capacity and more than 5 GW of private-grid delivery, initially outside ERCOT. For infrastructure buyers following Texas’s new data-center audit, the project puts the trade plainly: bypassing the interconnection queue can accelerate capacity, but it transfers fuel, emissions, and permitting exposure onto the campus itself.
Wire

Roku adds a 24/7 AI-generated TV channel
08/08/26, 04:31 PM

Roku added one 24/7 channel devoted to AI-generated programming, Fairground AI Creator TV, as part of a four-channel expansion of its free ad-supported streaming lineup. The Verge’s review of the launch says viewers cannot choose individual videos and Fairground does not disclose which generation models or training rights sit behind the feed, though it says creators receive generation payments plus an undisclosed revenue share. For media builders tracking the collapsing production barrier for AI video, the distribution milestone matters more than the programming: synthetic video has reached a living-room channel, but provenance, retention, and advertiser tolerance remain the tests before it becomes a durable format.
Wire

Android Canary exposes eight AI clock slots
08/08/26, 04:29 PM

Android Canary 2608 exposes an experimental Pixel interface for generating 10 numeral designs, from 0 through 9, rather than choosing only among the phone’s eight existing lockscreen clocks. 9to5Google’s teardown found a prompt box, suggested ideas, a reference-image upload, and a gallery for saved designs, but the server-side generator is not live and Canary features can disappear before release. For consumer-app teams tracking Google’s attempt to make Gemini the default layer across Android, the small signal is worth filing away: generation is moving from destination apps into operating-system personalization, where builders should wait for a public API and provenance behavior before planning an integration.
Wire

Spatial training lifts a VLM 27.63 points
08/08/26, 03:26 PM

Fine-tuning Qwen3-VL-8B on GST-Train raised its global-spatial-awareness score from 25.89 to 53.52, a 27.63-point gain, in the GST-Bench paper. The 2,762-question benchmark forces 22 vision-language models to reconstruct long video trajectories from novel viewpoints; the strongest zero-shot system scored 42.68 against 79.08 for humans. Robotics teams should file the result beside World Labs’ simulation-first robot-testing funnel: targeted spatial curricula may beat a larger zero-shot model, but synthetic-video gains still need validation against noisy physical sensors.