# Review Gates, Intent Files, and Cursor's Agent-Native Git

*By Coding Agents Alpha Tracker • June 17, 2026*

Review guardrails dominated today: Addy Osmani's data-rich case for risk-tiered agent PR review was the clearest signal. Also worth your time: Kent C. Dodds's INTENT.md pattern, Cursor's Origin launch, LangSmith Sandboxes, and a few concrete Cursor workflows.

## 🔥 TOP SIGNAL

- Addy Osmani's `Agentic Code Review` is the clearest practical read of the day: GitClear data says daily AI users generate about **4x** the code for only about **12%** more delivered value, while incidents-to-PR rose **242.7%**, per-developer defects went from **9% to 54%**, and median review duration rose **441.5%** [^1]. His answer is not to back away from agents, but to review differently: batch-triage PRs with Claude Code or Codex, use heterogeneous reviewers, tier by blast radius, and keep the merge decision human-owned [^1].

> "Treat CI as the wall that does not move." [^1]

## ⚡ TRY THIS

- **Turn review into a gated pipeline, not a vibe check.** Addy Osmani's playbook is straightforward: 1) point Claude Code or Codex at a batch of PRs and bucket them into `safe to merge`, `needs work`, and `high-risk`; 2) run **two different** AI reviewers on risky diffs; 3) tier depth by blast radius; 4) refuse review without an intent statement, test output, and a small diff; 5) read rewritten tests first, keep deterministic CI strict, and let a human own merge [^1].

- **Shift human effort into the plan, then automate the line-by-line gate.** Kun Chen's solo workflow: write a detailed plan up front, run **20-30 agents in parallel** for hours, stay on escalation for stuck agents, and gate merges through an automated `No Mistakes` review step. The transferable pattern is simple: human-owned intent before execution, automated verification after execution [^1].

- **Add an `INTENT.md` contract to long-lived packages.** Kent C. Dodds has Kody create and maintain an `INTENT.md` file describing package goals, then compare every proposed change against it. If the goal itself changed, the agent should only update `INTENT.md` when the user explicitly wants that change [^2].

- **Build internal tools your agent can operate, not just generate.** Riley Brown's Cursor demo prompt was essentially `make a to do app for me as a creator with a full database... be able to write to this database... make it look like a simple version of Notion, but dark mode`. He used Convex for the DB, let the agent add tasks by natural language, and deployed it with `@vercel put this on the internet` [^3].

## 📡 WHAT SHIPPED

- **Cursor Origin** — Cursor is launching code storage and git hosting so teams and agents can host, review, and collaborate on code; swyx/Tomas Reimers highlighted agent-specific features: scalability for agent workloads, API/MCP extensibility, built-in merge conflict resolution, and CI/CD failure resolution. Available this fall; [waitlist](https://cursor.com/origin-waitlist) [^4][^5][^6].

- **Cursor at Compile** — Michael Truell said **>95%** of Cursor users now use it primarily as an agent, and agent requests are used about **5x** more than assistive features. He also described Cursor 3 capabilities around gesture-based design edits, recursive sub-agents, days-long remote project handoffs, and broader SDK/CLI/plugin extensibility [^7].

- **AI reviewer comparison got sharper** — CodeRabbit topped the Martian benchmark on F1; Greptile was cited at about **82%** bug-catch versus CodeRabbit's **44%** in one benchmark; Anthropic said its internal Code Review had **<1%** incorrect findings and raised substantive reviews from **16% to 54%**. The operational takeaway from Addy's roundup: reviewer diversity matters, because in one 146-PR test **93.4%** of flagged locations were unique to a single tool [^1].

- **LangSmith Sandboxes** — LangChain positioned this as the right layer when an agent needs to *do* something: verify generated code runs before responding, operate on real files, persist state across tool calls, scale bursty parallel evals/RL, or safely handle user input that may be executed. [Blog](https://www.langchain.com/blog/give-your-ai-agent-its-own-computer) [^8][^9].

- **GLM 5.2 in Cursor via OpenRouter** — Riley Brown shared the exact setup: paste an OpenRouter key into Cursor's OpenAI API override, set the base URL to `https://openrouter.ai/api/v1`, then add custom model `z-ai/glm-5.2`. Context from Kalo: people he trusts were reporting strong results from GLM 5.2 [^10][^11].

## 🎬 GO DEEPER

- **12:00-14:30 — Riley Brown's agent-writable internal app demo.** Good clip if you want a concrete pattern instead of a slogan: prompt the app into existence, attach a database, let the agent write into it, then verify the state persists [^3].


[![SpaceX Just Bought Cursor for $60B. It’s About to Take OVER.](https://img.youtube.com/vi/LSpEP9N_7iY/hqdefault.jpg)](https://youtube.com/watch?v=LSpEP9N_7iY&t=720)
*SpaceX Just Bought Cursor for $60B. It’s About to Take OVER. (12:00)*


- **16:39-18:52 — Codex/Claude -> Cursor skills handoff.** Watch this if tool-switching friction is your blocker: Riley exports skills and memory into a `Codex Import` folder with a README and `Needed Keys`, then asks Cursor to import it globally [^3].


[![SpaceX Just Bought Cursor for $60B. It’s About to Take OVER.](https://img.youtube.com/vi/LSpEP9N_7iY/hqdefault.jpg)](https://youtube.com/watch?v=LSpEP9N_7iY&t=999)
*SpaceX Just Bought Cursor for $60B. It’s About to Take OVER. (16:39)*


- **8:13-9:11 — Michael Truell on the next agent handoff shape.** Short but high-signal: the target state is not three local agents for 30 minutes, but handing out whole projects and getting back completed, tested work days later [^7].


[![Cursor’s first Compile user conference: Michael Truell’s opening keynote](https://img.youtube.com/vi/-Exfie3VepY/hqdefault.jpg)](https://youtube.com/watch?v=-Exfie3VepY&t=493)
*Cursor’s first Compile user conference: Michael Truell’s opening keynote (8:13)*


- **Repo/file to study: `llama.cpp`'s `.pi/gg/SYSTEM.md`.** Georgi Gerganov's local setup is intentionally tiny—`pi -nc --offline` plus a short system prompt. Start with the [SYSTEM.md](https://github.com/ggml-org/llama.cpp/blob/master/.pi/gg/SYSTEM.md) and the [ggml-org Assisted-by commit trail](https://github.com/search?q=%22Assisted-by%22+user%3Aggml-org&type=commits&ref=advsearch) if you want a minimal maintainer-grade local-agent workflow [^12].

*Editorial take: more code is already cheap; the leverage has moved to intent control, review gates, and agent-native infrastructure.* [^1][^2][^1][^4]

---

### Sources

[^1]: [Agentic Code Review](https://addyo.substack.com/p/agentic-code-review)
[^2]: [𝕏 post by @kentcdodds](https://x.com/kentcdodds/status/2067005035062702416)
[^3]: [SpaceX Just Bought Cursor for $60B. It’s About to Take OVER.](https://www.youtube.com/watch?v=LSpEP9N_7iY)
[^4]: [𝕏 post by @cursor_ai](https://x.com/cursor_ai/status/2067012220832329782)
[^5]: [𝕏 post by @swyx](https://x.com/swyx/status/2066928345246470204)
[^6]: [𝕏 post by @swyx](https://x.com/swyx/status/2066936698848838050)
[^7]: [Cursor’s first Compile user conference: Michael Truell’s opening keynote](https://www.youtube.com/watch?v=-Exfie3VepY)
[^8]: [𝕏 post by @LangChain](https://x.com/LangChain/status/2066938607886012511)
[^9]: [𝕏 post by @LangChain](https://x.com/LangChain/status/2066938608750084299)
[^10]: [𝕏 post by @rileybrown](https://x.com/rileybrown/status/2067075406553895342)
[^11]: [𝕏 post by @kalomaze](https://x.com/kalomaze/status/2067013372697329735)
[^12]: [Quoting Georgi Gerganov](https://simonwillison.net/2026/Jun/16/georgi-gerganov)