Trending AI Briefing: Friday, November 20, 2026 (afternoon ET)
Friday afternoon, and the loudest AI headlines are about stock prices. The more useful story sits underneath them: the gains builders can actually bank this week came from the layer around the model, the harness that hands it tools and reads back its results. Hugging Face trained adapters for three specific coding harnesses and cut their tool calls by more than half, Microsoft Foundry let a browser agent batch a whole click sequence into one script, and GitHub, which Microsoft owns, let team maintainers set default agent plugins for everyone on a team. Same models, different plumbing, better results.
What's hottest in AI news right now
Hugging Face published "Introducing harness adapters" on November 20, a recipe for training a small add-on to an open model so it works better inside one particular agent harness. The adapter is a LoRA (a small set of extra weights on a frozen base model) trained with reinforcement learning on Qwen3.8-27B while the harness ran unmodified, with Pi 1.5.2, OpenCode 2.2.7 and Hermes Agent v0.22.6 as the three targets. On the hardest 51 SWE-bench Pro tasks, the post's results table shows the Pi adapter going from 40.0% to 64.0% with 60% fewer tool calls (its prose gives a slightly different 39.2% to 64.7%); across harnesses, tool calls fell 56% to 72% on SWE-bench Pro. These are the authors' own evals at one attempt per task. The catch, in their words: the merge weight for the combined "super-adapter" was chosen on Terminal-Bench 2.1, so its Terminal-Bench numbers are "slightly optimistic," and "you cannot compare the results across different harnesses." Adapters, GGUFs and the training tasks sit in the harness-hub collection. Hugging Face
Microsoft announced on November 18 that the Browser Automation tool in Foundry Agent Service, built on Azure Playwright Workspaces, is generally available. The headline feature is action batching: instead of a long chain of snapshot, click and type calls, an agent can send one Playwright script for a whole sequence through a run-code tool. Live View and Take control (a person can grab the browser mid-run to sign in, then hand it back) and opt-in session recording are also GA. VNet injection, bring-your-own proxy, CAPTCHA handling through Web Bot Auth and Key Vault secrets are still preview, and profile reuse is listed as coming soon. Microsoft keeps the run-code tools opt-in "because scripts can run any code," and the post gives no pricing. Microsoft Foundry blog
GitHub added eight GitHub-verified MCP connectors to Copilot on November 18, in public preview for all Copilot plans. The new services are Box, Buildkite, Calendly, ClickUp, Mixpanel, Supabase, Webflow and Wix, joining more than 25 already connectable. In Copilot CLI you open /mcp, pick the Connectors tab and select Connect. Business and Enterprise organizations need the "Editor preview features" policy on, and any organization running a per-connector MCP allowlist has to add the new ones by hand. GitHub changelog
GitHub also made team customizations generally available on November 18, letting team maintainers and org owners set approved Agent Plugins as defaults for a team. The defaults switch on for direct team members in VS Code, the Copilot app and Copilot CLI on Business and Enterprise plans. Child teams do not inherit them, members can turn a plugin off where policy allows, and enterprise or org policy overrides a team's choice. The starting plugin list for a team's agents is now something a maintainer sets, inside limits the org controls. GitHub changelog
A paper titled "Language Models Can Rewrite Their Own Weights" reached about 240 points on Hacker News today, four days after its November 16 arXiv submission. The authors give a 27B model its own weights, graded answers and the ability to run itself, and report that its self-made edit, with no gradients or training, raised accuracy by 6.6 points on 490 unseen problems. They say the ability only appears at that size: their 2B model made no edit, and the 4B and 9B models "break themselves." It is one lab's preprint, not yet peer reviewed. arXiv
An OpenJDK pull request that a Hacker News commenter ties to Java4Shell was opened on September 9 and closed unmerged on November 4. PR #32781, titled "AVX-512 Base64 decoder overwrites output buffer on short padded tail," describes a stub that "stores past the logical output length... corrupting the heap" on invalid input, with a one-line early-return fix. The author wrote on November 20 that the work was "taken over by others" and the issue "is gone on master." The page never names CVE-2026-114938; the Java4Shell link comes from a commenter, though the description matches the bug Calif disclosed this week. If the link holds, the fix was publicly visible for about ten weeks before Calif's disclosure. OpenJDK PR #32781
New tools and features worth actually trying
Codex CLI 0.170.0. Released November 18 (npm install -g @openai/codex@0.170.0), it fixes MCP OAuth races, adds experimental support for identically named MCP servers across environments, and falls back to HTTP when a WebSocket request is oversized. Honest tradeoff: the cross-environment MCP feature is marked experimental, so do not build a team setup around it yet. ChatGPT and Codex changelog
gh event listen for agents that wait on pull requests. GitHub CLI 2.105.0 can block until a PR event arrives (gh event listen <PR URL> with --count, --follow, --timeout), so an agent stops polling for review comments or CI results. Honest tradeoff: the commands are preview and depend on Event Streams, which is itself in preview behind a waitlist. GitHub changelog
Gemini 4 Neon in GitHub Copilot. Generally available since November 18 for Copilot Pro+, Max, Business and Enterprise, rolling out gradually, and admins gate it through model policy. Honest tradeoff: it bills at provider list prices under usage-based billing, so a team that switches defaults without watching spend will find out on the invoice. GitHub changelog
The harness adapters themselves. If you already run Pi, OpenCode or Hermes Agent on a local Qwen3.8-27B, the published specialist adapter for your harness is the cheapest experiment on this list. Honest tradeoff: it only helps that base model and that harness, and the gains are the authors' own measurements. Hugging Face
Trending AI repos on GitHub today
Read from Trendshift at about 15:15 ET; its rankings are momentum scores, and star counts below come from cache-busted shields badges, which round.
- simplexai-labs/linger (#2): a personal agent that runs a local model on your own machine and files your mail, chats and meetings into searchable wiki pages each night. Why now: the repo only dates to November 10, and its desktop tags have moved from v0.1.3 to v0.1.6 between November 18 and November 20. Star figures conflicted between sources, so none is given. License is a modified Apache-2.0 that requires a commercial license to offer Linger as a hosted service; its speed and memory scores are the company's own, some measured on simulated GPU tiers.
- mdiri-ai/SoduWork (#1): a self-hosted engine that dispatches work across agent harnesses, has a second agent review it and asks a person to sign off. Why now: top of the board again, with tags up to v1.13.19. MIT; star figures conflicted between sources. Bring your own model keys.
- HKUDS/CLI-Anything (#17): generates JSON-speaking command-line wrappers so agents can drive desktop software like GIMP or LibreOffice, plus a registry to install them (
pip install cli-anything-hub). Why now: back on the board. Apache-2.0, about 60k stars, latest tag v0.4.0 from June. The wrapped apps still have to be installed separately. - zhouxiaoka/autoclip (#7): a desktop, CLI and MCP tool that cuts a long video into captioned short clips with covers and posting copy. Why now: v1.5.8 in early November. MIT, roughly 11k to 12k stars depending on source. The cloud analysis models bill per use on your own key and receive transcripts and sampled frames; a local Ollama model is the private option.
- iamajmeena/neonclipping (#8): a local tool that scores audio loudness to pick moments from long streams and turns them into 9:16 captioned clips. Why now: v1.0.0 tagged November 19. MIT. The two-minute speed claim is the author's own; it needs Python 3.10+ and FFmpeg.
- cathrynlavery/diagram-design (#25): an agent skill for Claude Code, Codex and others that draws editorial-style diagrams as one self-contained HTML and SVG file. Why now: new on the daily board. MIT, roughly 60k to 69k stars depending on source. No tagged releases; versions live in a changelog.
- BaaRai/ARTEX (#18): a self-hosted autonomous penetration-testing platform where a planner splits goals and workers run real tools against targets. Why now: fast momentum. AGPL-3.0, with git tags up to v0.6.0 but no published GitHub releases. It is an English translation of an upstream repo the README says "is no longer public," so provenance is thin, and it is an offensive tool that needs written scope.
What actually matters from today's signal
Track the harness. The model leaderboards still get the headlines (a community "Models Engineering Index" put Gemini 4 Neon at #1 overnight), but this week's measurable wins came from changing what surrounds the model: an adapter trained against one harness's exact token stream, a browser tool that accepts a script instead of forty clicks, a plugin list a maintainer sets once for a whole team. For builders the highest-signal areas are local open models tuned to the harness you already use, batched actions for browser and desktop agents, team-level plugin defaults and verified MCP connectors, and event-driven agents that wait instead of poll.
The counter-signal is that every one of these moves concentrates trust in a layer fewer people inspect. Microsoft itself fences run-code behind an opt-in. A team default means one maintainer's plugin choice reaches every direct member unless they switch it off. And control cuts both ways: a post that drew about 590 points on Hacker News since Thursday, "Fuck your AI safeguards," describes a developer blocked by Claude Code's cyber safeguards from fixing his own service, who then found the fix with an open-weights model, which is exactly the harness-level escape hatch these platforms cannot close. The OpenJDK pull request is the sharper warning, if the commenter's link to Java4Shell holds. A fix for a heap-corruption bug that is publicly visible for weeks gives anyone with an agent that reads patches faster than vendors ship advisories a head start on you.
Source access notes: Read Friday November 20 from about 15:10 ET. The workspace shell reached only npm and raw GitHub; vendor pages, Hacker News (via the Algolia API) and Trendshift were read through web-fetch subagents, and release data came from GitHub release pages because the releases.atom feeds were refused, then tags were re-checked with git ls-remote: the releases pages lagged the tags for linger, SoduWork and ARTEX, and the repo lines use the tags. openai.com/news failed at its site; Google's AI blog served a stale snippet; DeepMind shows month-level dates only; the Codex changelog now redirects to learn.chatgpt.com, which was read there; Mistral and xAI had nothing in the window; Product Hunt was skipped. Anthropic's news page shows a November 20 "Announcements" label with no readable title, so nothing is reported from it. Claude Code 2.1.330 reached npm at 19:16 UTC today on the next tag, but the changelog on main still ends at 2.1.329, so it is not covered. Anthropic's Nasdaq debut and the market slide dominated Hacker News but carry no builder angle and the pricing ran in yesterday's brief. Stories and repos from the November 20 morning and November 19 afternoon briefings were skipped. A Sonnet 5.5 adversarial pass then ran on the draft. Applied: limited the Hugging Face "slightly optimistic" caveat to the super-adapter's Terminal-Bench numbers (it had sat next to SWE-bench Pro figures) and flagged the post's prose-versus-table mismatch, narrowed the GitHub claims to team-default plugins (connectors are a separate org-level preview, and members can switch defaults off), softened the OpenJDK closing to keep the commenter's Java4Shell link conditional, and replaced star counts where shields and repo pages disagreed. Held: the safeguards post description, which comes from a direct read of the post. Codex's fact checks on the two articles built from this brief confirmed the Hugging Face and Linger figures used here and changed no briefing claim.