Saturday Tools: The Editor Disappears

Each Saturday, The Wise Operator digs into the tools, tutorials, and trending builds worth your weekend. No news, just hands-on discovery.

It’s Saturday. The news can wait. This week three releases land in the same direction. Cursor shipped its own agentic-coding model. Google relaunched Antigravity as a five-surface agent platform with a brand new CLI. Anthropic walked into London and added sandboxes and tunnels that let Claude agents run on a company’s own infrastructure without ever touching the public internet. The editor used to be where the operator did the work. This week the editor became the room the operator watches the work get done in.

Tools Worth Your Weekend

Cursor Composer 2.5

Cursor shipped its first in-house agentic coding model on May 18, built on Kimi K2.5 and trained on roughly twenty-five times more synthetic tool-use traces than the prior version. Composer 2.5 scores 79.8 percent on SWE-Bench Multilingual, essentially tying Claude Opus 4.7 at 80.5 percent, and edges Opus 4.7 on Cursor’s own CursorBench v3.1 at 63.2 versus 61.6 percent. Standard tier runs $0.50 per million input tokens and $2.50 per million output, roughly one-tenth Opus pricing. The fast variant is $3 input and $15 output. Included usage doubles through May 25 as a launch promo.

Why it matters: Cursor is no longer a wrapper around someone else’s frontier model. It is a coding agent with its own model, tuned for the verbs the editor actually uses: file edits, terminal calls, test runs, and long tool-use loops. The bet is that an in-house model trained for one editor’s habits beats a general-purpose frontier model routed through the same editor. The benchmarks say the bet works for long sessions. The deeper signal is that the frontier-lab-as-supplier era is starting to crack at the edges.

Try this: Open a side project that is two or three commits past where you remember the architecture. Hand the whole repo to Composer 2.5 at default effort and ask it to write an architectural audit, run the test suite, and propose a refactor. Watch how many tool calls it chains before pausing for you. That count tells you whether you are paying for an editor or for an agent loop cost that compounded faster than you noticed. Paid; Pro plan starts at $20 per month.

Google Antigravity 2.0

Google relaunched Antigravity at I/O 2026 on May 19 as five surfaces under one roof: a standalone desktop IDE, a brand new Go-based CLI that replaces the old Gemini CLI, an SDK, a Managed Agents tier inside the Gemini API, and an Enterprise Agent Platform for big-company deployments. The platform runs on Gemini 3.5 Flash by default and adds dynamic subagent orchestration, scheduled background tasks, and native voice commands. A new $100 AI Ultra tier sits below the existing top plan, which dropped from $250 to $200, and grants five times higher Antigravity limits than Pro.

Why it matters: Google is no longer asking what an IDE plus AI should look like. It is asking what the developer’s entire working surface should be when the agent is the primary worker. The Go CLI signals where Google expects heavy users to live: the terminal, not a window. The Managed Agents tier means Google will run your agent in their cloud, on a schedule, against your repo, and you check on it through the IDE or the CLI. That is the same architectural turn Anthropic made with Routines five weeks ago, now mirrored by the company that owns Android and the search box.

Try this: Install the Antigravity CLI alongside your existing Claude Code or Cursor setup. Point it at the same repo you used Composer 2.5 against above. Give both the identical prompt: “audit the architecture and propose three refactors with diffs.” Compare the diffs side by side, not for which is better, but for which vocabulary each agent uses to describe the same code. The vocabulary gap tells you what each model was actually trained to optimize for. Free preview tier is tightly capped; serious use needs a paid plan.

Claude Agent Sandboxes and MCP Tunnels

At Code with Claude London on May 20 and 21, Anthropic added two pieces of agent infrastructure that quietly change who can run a Claude agent in production. Agent sandboxes let a company host the agent’s execution environment on its own servers instead of on Anthropic’s. MCP tunnels let that on-premise agent reach internal databases, ticketing systems, and document stores without exposing any of them to the public internet. Both ship as part of Claude Managed Agents, alongside the multi-agent orchestration, outcomes, and “dreaming” capabilities Anthropic introduced in San Francisco two weeks earlier.

Why it matters: the previous version of running Claude inside an enterprise required piping company data through Anthropic’s cloud or routing it through a separate non-human-identity gateway. Sandboxes and tunnels remove that tradeoff. A legal team can now run a Claude agent that reads from the internal contracts database without that database ever appearing on a public endpoint. The “where does the work run” question, which TWO has been circling for months, finally has an enterprise-grade answer that does not require the operator to forfeit data control.

Try this: If you have a side project with any sensitive data, draft a one-page architecture diagram for how you would deploy the same agent twice, once on the public Managed Agents tier and once via a sandbox plus tunnel. The exercise forces you to name which data crosses which boundary. That naming is the work that the convenience of a chat tab let you skip for the last two years. Available on Claude Team and Enterprise plans.

What to Watch

Google Antigravity 2.0 Beginner’s Guide walks through the new desktop app, the Go CLI, and the first dynamic subagent run, with the IDE side by side with the terminal so you see what each surface actually shows. Twenty minutes, no hype, no install pad. Watch the segment where the host hands the agent a multi-file refactor and switches to the CLI mid-run. That switch is the move most new Antigravity users miss, and it is what makes the platform feel like one tool instead of three.

Wisdom Speaks

“The man of practical wisdom is the one who is able to deliberate well about what is good and expedient for himself.” Aristotle, Nicomachean Ethics, VI.5

Cursor’s in-house model, Antigravity’s subagent fleet, and Claude’s sandboxed agent each enlarge what one operator can execute in a Saturday afternoon. None of them enlarge her capacity to decide which Saturday afternoon is worth executing in. Aristotle named the gap two thousand years ago. Techne is the craft of making the thing well. Phronesis is the older, slower wisdom of knowing whether the thing should be made at all. The agents arriving this week shift techne from the operator’s hands into the platform’s loop. Phronesis stays where it always was.

Solomon’s prayer in 1 Kings 3:9 asks for one specific gift. “Give therefore thy servant an understanding heart to judge thy people, that I may discern between good and bad” (KJV). He could have asked for armies, for wealth, for execution speed. He asked for discernment, because he already knew the throne would amplify whatever heart sat on it. The agentic IDE is the throne the operator inherits this week. The question Solomon asked is the same one. The execution capacity is granted. The discerning heart is still the work that is hers.

Last Saturday in this seat: Saturday Tools: Daybreak, Claude for Small Business, and OpenHuman Leave the Chat Tab, on the week the agent left the chat tab. This Saturday the editor follows it out the door.