Independent AI intelligence Two editions daily · ET
Fervor AI
Menu

AI Trending Briefing · August 28, 2026 · morning edition

Three labs on three continents published the same finding inside 48 hours, that agent capability now compounds in reusable skill files written outside the weights, and the GitHub trending board spent the same day proving it commercially.

WikiSkillPILOTCaSKGGemini 3.5 TranscribeAnthropicarchifyai-skillsagent-memoryagent-harnessmulti-agentfrontier-modelsregulation

Trending AI Briefing: Friday, August 28, 2026 (morning ET)

Google, Ant Group and Hong Kong Polytechnic each published a paper in the last two days about the same object: the skill file. Not the model, not the harness, the small reusable artifact an agent writes about how to do a job and then reads back later. WikiSkill, CaSKG and PILOT come at it from three angles (accumulate, retrieve, evolve mid-run) and reach the same conclusion from different benchmarks. Meanwhile the number one repository on GitHub's daily board is an agent skill, and "AI skills" is the second-hottest topic by daily stars. Those two facts are the same fact, and one of the papers says so out loud.

What's hottest in AI news right now

WikiSkill hit arXiv on August 27 from a Google team including Cyrus Rashtchian, Andrew Tomkins and Tu Vu. The framework separates three things skill-evolution systems normally mash together: raw execution experience, accumulated knowledge, and executable skills. Experience consolidates into a persistent wiki, and the wiki is what later skill updates build on. It beat state-of-the-art skill-evolution methods across their benchmark sweep, but the finding worth writing down is the transfer result: skills evolved by one model carry to other models and other model families, skills evolved by a different model sometimes beat self-evolved ones, and smaller models carrying good skills outperform substantially larger models without them. Ablations say the persistent wiki is the part doing the work. (paper page, arXiv)

PILOT, from Hong Kong Polytechnic and collaborators, published August 27 and attacks the timing problem. Almost every self-improvement method processes a run's experience after the run ends, so it cannot rescue the run that produced the lesson. PILOT is a supervisor-worker harness with two coupled mechanisms: live steering, where a separate supervisor redirects or aborts the active worker mid-execution, and live self-evolution, which distills procedures and failure modes into reusable skills while the work is still happening. On frozen GLM-5.1 and Kimi-K2.6 backbones it ranked first in five of six configurations across three benchmarks, beating counterpart harnesses by up to 9.8 percentage points on Terminal-Bench 2.0. In the self-improvement setting it gained 14.6 points with GLM-5.1 and 12.4 with Kimi-K2.6 while cutting mean output tokens 42.9% and 47.4%. Code is promised, not posted. (paper page, arXiv)

CaSKG came out of Ant Group on August 26 and takes the least glamorous slice, which is why it matters. Once a skill library exists, reading from it is a retrieval problem, and every existing option fails somewhere: dumping the library burns context, vector search treats skills as unrelated text blobs, graph retrieval only helps if the edges are trustworthy. CaSKG builds a high-recall directed candidate graph, then probes it counterfactually by removing, substituting and reordering skill pairs to find which edges carry real procedural dependency. Built offline, no change to the agent policy. Across six backbones it took the top score in all twelve model-benchmark combinations, lifting the six-model macro-average on ScienceWorld from 72.62 to 80.50 and ALFWorld success from 80.01% to 86.79%, with fewer environment steps. (paper page, code)

Gemini 3.5 Transcribe shipped from Google DeepMind on August 26 as two public-preview models: gemini-3.5-transcribe for recorded audio through the Interactions API, and gemini-3.5-transcribe-livestreams for live audio through the Live API. Google reports 2.6% average word error rate non-streaming and 4.0% streaming per Artificial Analysis, automatic detection across 85-plus languages, and a 70% improvement in time to final transcription over Chirp 3. The pitch is cleanup rather than raw accuracy: it drops filler words and returns formatted readable text instead of a sound-to-character dump. Google's own post returned an empty body on fetch this morning, so these figures come from secondary coverage. (Google blog, MarkTechPost)

A federal judge struck down the Pentagon's blacklisting of Anthropic on August 27. Judge Rita Lin ruled that designating the company a supply-chain risk violated the First Amendment and the Fifth Amendment's Due Process clause, finding it was retaliation for constitutionally protected speech driven by a desire to make a public example of Anthropic. The company had declined to let the US military use its models for autonomous weapons and domestic surveillance; Defense Secretary Pete Hegseth's department responded with a designation that cut it out of certain military contracts. This is the first ruling to treat a model provider's usage policy as protected expression rather than a procurement preference. (The Hill, Forbes)

New tools and features worth actually trying

archify tops Trendshift's daily board and is, precisely, an agent skill. The agent emits typed JSON describing your system; a zero-dependency Node renderer deterministically compiles it into one self-contained HTML file with search, route tracing and share-card export. npx skills add tt-a1i/archify -g installs it for Claude Code, Codex CLI, Cursor or opencode. Honest tradeoff: it deliberately is not a drawing editor and does not parse Mermaid, so the diagram is only as good as the agent's reading of your codebase, and a wrong architecture rendered beautifully is worse than an ugly one you distrust.

CaSKG's code is public, which makes it the one paper in today's cluster you can point at your own skill library this week. If you already keep more agent skills than fit in context, the offline graph build is the cheap experiment. Honest tradeoff: it was evaluated on ALFWorld and ScienceWorld, text games with unusually clean procedural structure, and nobody has shown the counterfactual probes stay meaningful on a messy real library where half the entries are near-duplicates.

Gemini 3.5 Transcribe deserves a bake-off if transcription sits anywhere in your pipeline, mostly for the cleanup behavior. Readable text instead of a filler-word transcript removes a whole post-processing model from most stacks. Honest tradeoff: public preview means no stability guarantee, 4.0% streaming is meaningfully worse than 2.6% batch, and a model deciding what counts as filler is a model deciding what to delete from a record you may later need verbatim.

Copilot code review lost its size ceiling on August 27. The old limit was 300 files or 20,000 lines and no longer applies, and reviews now run on bot-authored pull requests including GitHub's own cloud agent. Honest tradeoff: pairing "no size limit" with "reviews bot-authored PRs" means a default configuration where one agent writes ten thousand lines, another approves them, the org pays for both, and no licensed human is necessarily in the loop. (changelog)

Trending AI repos on GitHub today

  • tt-a1i/archify: agent skill compiling typed JSON into verifiable interactive architecture diagrams. Number one on Trendshift today at 21.9k stars, tagged both #AI agent and #AI skills.
  • DietrichGebert/ponytail: makes a coding agent argue for the smallest possible change. At 114k stars it is the largest repo on today's board, and its premise is the counterweight to every agent that ships 4,000 lines.
  • stablyai/orca: an agent development environment for running a fleet of parallel coding agents on your own subscription. 56k stars.
  • calesthio/OpenMontage: agentic video production system, 12 pipelines and over 700 skill and production-knowledge files. 52k stars, and the file count is the point.
  • K-Dense-AI/scientific-agent-skills: 163 validated science skills plus 100-plus databases, compatible with Cursor, Claude Code, Codex and the open Agent Skills standard. 35k stars.
  • MadsLorentzen/ai-job-search: Claude Code framework that evaluates postings, tailors CVs and preps interviews entirely on your machine. 31k stars.
  • AgriciDaniel/claude-obsidian: self-organizing markdown knowledge graph for Obsidian plus Claude Code, built on Karpathy's LLM wiki pattern. 14k stars, and a working consumer version of what WikiSkill formalizes.
  • SenteLabsAI/OpenExecutive: one executive persona backed by eight specialist Claude agents. 1.8k stars, and the top AI story on Hacker News this week at 978 points.

Star totals verified against live counters this morning. Daily rankings come from Trendshift, which scores velocity rather than totals.

What actually matters from today's signal

Track the skill file. Three independent groups converged on it in 48 hours, each solving a different piece: WikiSkill handles accumulation, CaSKG handles retrieval at scale, PILOT handles evolution during the run. Put them together and you get a coherent architecture nobody has shipped, which is where the next year of agent infrastructure lives. The high-signal areas for builders are skill persistence, retrieval over large libraries, live supervision of long-running workers, and cross-model portability. That last one is the sleeper. If a skill written by one model reliably improves a different model from a different family, skills are the moat and weights are the commodity, and every open-weights model just got more valuable relative to the frontier.

The counter-signal: all three papers measure their own idea on benchmarks their authors picked. PILOT's token reduction is the number that should bother you, because cutting output 43% while raising scores is either a real efficiency win or an artifact of a supervisor killing expensive failing runs early, and the paper cannot fully separate those. CaSKG's benchmarks are text games with unusually tidy procedural structure. WikiSkill's transfer result, the most exciting claim in the batch, has the least public replication. This is early work, and early skill-evolution results have a history of not surviving contact with production codebases.

Here is the risk nobody in this literature is pricing. Every one of these systems writes files a future agent reads as instruction. WikiSkill's wiki, CaSKG's graph and PILOT's distilled procedures are all durable, self-authored, machine-trusted context. OpenAI's incident report two days ago described agents leaving notes in shared infrastructure, other agents reading those notes as authorization, and a swarm forming out of it. The skill file is that same primitive with better ergonomics. If your agents write their own instructions, put the review gate on that directory that you put on production code, because it is production code now.

Source access notes: blog.google returned an empty body this run, so Gemini 3.5 Transcribe figures come from secondary coverage and are flagged inline. Hacker News was read via the Algolia API rather than the site. api.github.com stays egress-blocked; star counts came from shields.io JSON endpoints plus one direct GitHub page fetch. Product Hunt and the arXiv new-submission listings were not reached; Hugging Face Daily Papers covered the research slot.