Independent AI intelligence Two editions daily · ET
FervorAI
Menu

AI Trending Briefing · July 28, 2026 · afternoon edition

Agent capability now ships in two competing packages, and on the day MCP finalized a governed spec with deprecation policy and OAuth hardening, the trending board belonged to plain folders with a SKILL.md and no governance at all.

MCP 2026-07-28Agent SkillsAlibaba open-code-reviewOrcaPilot ProtocolEnterprise-Managed Authorizationmcpai-skillsagent-securityagent-infrastructuremulti-agentfunding

Trending AI Briefing: Tuesday, July 28, 2026 (afternoon ET)

Two ways of packaging agent capability collided on the same calendar day. MCP published the 2026-07-28 specification, its largest revision since launch, complete with a formal extensions framework, reverse-DNS identifiers, a twelve month deprecation policy, and mandatory OAuth 2.1 discovery. Twelve hours later the daily trending board was not an MCP board. Six of Trendshift's top twenty-five are agent skills: a folder, a SKILL.md, no signing, no sandbox, no review. Alibaba shipped its internal code review system with a skill install path. The #AI skills topic pulled roughly 13,000 stars today, second only to #AI agent. One format got a constitution this morning. The other one got the users.

What's hottest in AI news right now

MCP 2026-07-28 ships today as the final specification, ten weeks after the release candidate locked on May 21. The initialize handshake is gone, the Mcp-Session-Id header is gone, and protocol version plus client identity now ride in _meta on every request, with server/discover for on-demand capability lookup. A remote server that needed sticky routing, a shared session store, and deep packet inspection at the gateway can now sit behind a plain round-robin load balancer and route on the Mcp-Method and Mcp-Name headers. Extensions become first class with their own repositories and independent versions, MCP Apps and Tasks ship as the first two, and Roots, Sampling, and Logging are deprecated with July 28, 2027 as the earliest removal date. One caveat worth stating plainly: as of this afternoon the MCP blog's post index still tops out at the June 29 SDK betas announcement, so the ship date is confirmed by the release candidate timeline and secondary coverage rather than by a final post on the vendor blog. (MCP blog, AAIF migration guide, WorkOS)

Agent Skills, the specification Anthropic published at agentskills.io on December 18, 2025 and handed to the Agentic AI Foundation, is the format doing the actual distribution work this quarter. Roughly forty compatible products sit on the official showcase as of June 2026, including OpenAI Codex, GitHub Copilot, Cursor, Gemini CLI, and VS Code, and Atlassian, Cloudflare, Figma, Notion, Sentry, Stripe, and Zapier all publish through Anthropic's directory. Scale numbers come from community indexes rather than a registry of record, so treat them as directional: SkillsMP claims roughly 1.9 million public skills scraped from GitHub, and SkillsBench reports an average quality score of 6.2 out of 12 across 47,150 analyzed skills. (Agentman, Atlan)

alibaba/open-code-review is the clearest evidence that skills stopped being a hobbyist format. It is Alibaba Group's internal AI code review assistant, the one the company says served tens of thousands of developers and caught millions of defects over two years, now open sourced. The architecture is deliberately hybrid: deterministic pipelines plus an LLM agent, line-level comments, and a fine-tuned ruleset targeting null pointer exceptions, thread safety, XSS, and SQL injection. It reads Git diffs and hands changed files to a configurable OpenAI or Anthropic compatible model that can read full files and search the codebase for context. The install matrix includes GitHub Actions, GitLab CI, npm, and a skill at skills/open-code-review/SKILL.md. Fourteenth on today's Trendshift board. (GitHub, SKILL.md)

stablyai/orca is trending at eighteen and treats the agent itself as a fungible unit. It calls itself an ADE, an agent development environment, for running a fleet of coding agents in parallel: Claude Code, Codex CLI, OpenCode, Pi, Grok, and more than twenty others, each on your own subscription rather than a per-seat license. Every agent gets an isolated git worktree, so you can fire five at the same prompt and diff the results. It opens a worktree straight from a GitHub PR or Linear ticket with task context attached, and a mobile companion pings you when a run finishes. Secondary coverage puts it past 12,500 stars with Y Combinator backing as of July 2026; both figures are reported rather than verified. (GitHub)

Pilot Protocol left stealth on July 27 with a $4.5 million seed led by Version One Ventures, joined by Precursor Ventures, Night Capital, Todd & Rahul Capital, and angels including Lenny Rachitsky and Ben Tossell. The pitch is a discovery and trust layer where agents find each other, hand off work, and eventually pay each other in credits with no human wiring the integration. The company reports roughly 250,000 connected agents generating about two billion requests a day, and an app store where partner software got more than 30,000 autonomous installs in two weeks. Every one of those numbers is company-provided and independently unverified, which the coverage states explicitly. (Tech Startups, Pilot)

New tools and features worth actually trying

alibaba/open-code-review, this week, on a repository you know well enough to grade the output. The deterministic-pipeline-plus-agent split is the design most AI review tools skip, and the fine-tuned ruleset gives you a floor that does not depend on model mood. Honest tradeoff: the ruleset was tuned on Alibaba's codebase and skews toward Java and web-service failure modes, so a Rust or Go shop should budget real configuration time before the signal beats the noise. (GitHub)

tt-a1i/archify is for teams whose architecture diagrams are stale because keeping them current is nobody's job. One agent skill, working across Cursor, Claude Code, Codex CLI, and OpenCode, turns a system description or a repository into architecture, workflow, sequence, data-flow, and lifecycle diagrams exported as self-contained HTML plus PNG, SVG, and WebM. Honest tradeoff: a diagram generated from a repo reflects what the code says, not what the team meant, and a confident wrong diagram beats no diagram only until someone new treats it as ground truth. (GitHub)

mvanhorn/last30days-skill is worth thirty minutes for anyone who runs research passes as part of the job. It searches Reddit, Hacker News, Polymarket, GitHub, and the web in parallel with no configuration, adds X, YouTube, arXiv, and Techmeme through a setup wizard, then scores sources by engagement before an agent judge synthesizes one brief. Honest tradeoff: ranking by engagement means ranking by what got attention, which is a different quantity from what is true. (GitHub)

stablyai/orca if you have hit the wall where one agent plus one context window is the bottleneck. Five agents on isolated worktrees against the same prompt, diffed side by side, is a different workflow from prompting harder. Honest tradeoff: five parallel agents burn five times the tokens and produce five diffs a human still has to read, so the win only lands on problems where you cannot predict which approach works. (GitHub)

Trending AI repos on GitHub today

  • MoonshotAI/Kimi-K3: weights, report, and license for the 2.8T open frontier model. Holding first for a second day. (link)
  • kvcache-ai/AgentENV: distributed platform for agent environments on Firecracker microVMs, E2B-compatible. Second, and the most reusable piece of the Kimi drop. (link)
  • bradautomates/claude-video: a /watch skill that pulls a video from roughly 1,800 sites, extracts frames adaptively, and hands a timestamped transcript plus images to the model. Third, and proof of how small a useful skill can be. (link)
  • ayghri/i-have-adhd: a skill that rewrites agent output to lead with the answer and cut preambles. Ninth, and it changes nothing about capability, only presentation. (link)
  • alibaba/open-code-review: hybrid deterministic-plus-agent code review from Alibaba's internal tooling, with Actions, GitLab CI, and skill install paths. Fourteenth. (link)
  • earthtojake/text-to-cad: agent skills for CAD, robotics, and hardware design, producing STEP, STL, DXF, and robot structure files with browser previews. Seventeenth. (link)
  • stablyai/orca: ADE for running a fleet of coding agents in parallel on isolated worktrees. Eighteenth. (link)
  • virgiliojr94/book-to-skill: converts a technical book PDF into an agent skill you can reference while working. Twentieth. (link)

Ordering comes from the Trendshift daily board rather than API calls, so treat the ranking as the signal and any star count as approximate.

What actually matters from today's signal

The format war is the story, and it is not being fought on merit. MCP won the protocol argument and spent twenty months earning it: governance under the Linux Foundation, a SEP process, a deprecation policy, OAuth 2.1 and RFC 8707 resource indicators, extensions with delegated maintainers. Skills won the distribution argument by being a folder. The four highest-signal areas this week: the stateless MCP migration and specifically the Tasks lifecycle rewrite, Enterprise-Managed Authorization now stable and being adopted by Anthropic, Microsoft, and Okta, the skill as the install unit for enterprise tooling rather than a package manager, and parallel-agent harnesses like Orca that make the agent itself disposable.

The counter-signal is sitting in plain sight and almost nobody is pricing it. A February 2026 Snyk audit of 3,984 skills from the ClawHub registry found 36.82 percent carried at least one security flaw, 13.4 percent carried critical-severity issues, and 76 contained confirmed malicious payloads. The ClawHavoc campaign pushed more than 1,200 malicious skills into a marketplace between January and February. Publishing a skill requires a SKILL.md and a week-old GitHub account. No signing, no sandbox, no review. Compare that to what MCP shipped today, where a client must validate which authorization server issued a response and bind credentials to an issuer to close a mix-up attack class. The governed format hardened against confused-deputy attacks on the same day the ungoverned one took the top of the charts. (Snyk findings via Pluto Security, CSA research note)

Here is the move. Skills are winning because they are the right size, and that is not going to reverse, so stop waiting for it to. Treat a skill install exactly like a dependency install: pin the commit, read the SKILL.md before it reaches your context, and keep a list of what is loaded in every agent your team runs. The instructions inside a skill are not data, they are a program with your credentials, and the registry your team pulls from has less review than npm did in 2015.

Source access notes: WebFetch enforces a URL provenance rule, so every vendor page was reached through a search result first. blog.modelcontextprotocol.io, trendshift.io, workos.com, aaif.io, and techstartups.com all resolved. github.com/trending exceeded the response token limit, and direct fetches on individual github.com repository pages failed the provenance rule, so repos were verified through search-surfaced GitHub URLs. api.github.com was not called per standing instruction, so star counts are qualitative or attributed to secondary coverage. The Hacker News Firebase top-stories endpoint resolved but individual item URLs failed provenance, so HN was skipped. Product Hunt and arXiv new-submission listings were not fetched.