# Glasshand rents your agent a real iPhone, and Anthropic explains dreaming

> Agents on real phones, memory that corrects itself between sessions, voice calls in Cursor, and what else shipped, each with its source.

- Published: 2026-11-07
- Publisher: Rig Labs (https://riglabs.ai/)
- Canonical: https://riglabs.ai/briefs/2026-11-07/

---

## Glasshand

**What it is:** a service that rents dedicated physical iPhones and Mac minis in a data center to AI agents. Codapt Inc. runs it.

**What it does:** an agent such as Claude Code, Codex or Cursor connects over an API, sees the screen, and taps, types and clicks like a person. You can install App Store apps, sign in to your accounts, and add an eSIM plan to receive verification texts. [Glasshand's agent docs](https://glasshand.com/agents.md) list an iPhone SE (3rd generation) at $60 a month and a Mac mini M4 at $90 a month, plus tax, month to month. The API gives screen capture and input only, with no shell, file mount or clipboard.

**What to know first:** [the privacy policy says](https://glasshand.com/privacy) screen images can reveal anything visible on the device, and [the terms say](https://glasshand.com/terms) an authorized agent can act without a new approval for each action. Start with a throwaway account.

**Where it comes from:** Glasshand, launched on Product Hunt this week. [The site](https://glasshand.com).

## Dreaming, from Anthropic

**What it is:** an offline pass where an agent reads its past transcripts, corrects wrong notes, and builds a memory for later sessions. Anthropic explains it in an engineering post.

**What it does:** [Anthropic reports](https://www.anthropic.com/research/components-of-agent-memory) that a wiki built from about 350,000 files went from 13% to 48% of graded criteria on 25 questions as dreaming scaled from 2 sessions to 128. These are Anthropic's own evaluations. To try it, [the cookbook](https://platform.claude.com/cookbook/managed-agents-cma-dreaming-with-a-dynamic-workflow) builds a small version on Claude Managed Agents, at $0.36 on its example set and about $8 for fifty long transcripts. It warns that nothing reviews the result by default, so a bad transcript can reach your agent the next morning.

**Where it comes from:** Anthropic, post dated 2026-11-06.

## Voice chat with agents in Cursor

**What it is:** a realtime voice call with your coding agents, from Cursor, running on Grok Voice.

**What it does:** you describe a change out loud, ask what an agent is working on, or ask what changed in a diff, and hear progress as it works. It runs in the Agents Window, the IDE, the web and Cursor for iOS, where "Hey Siri, call Cursor" starts a call with a new agent. The page gives no price or limits.

**Where it comes from:** Cursor, 2026-11-06. [The changelog](https://cursor.com/changelog/voice-chat).

## Claude Code cloud sessions set up your repo

**What it is:** a change from Anthropic to Claude Code cloud sessions, for Pro, Max and Ultra plans.

**What it does:** [Anthropic says](https://x.com/ClaudeDevs/status/2118795192090386454) Claude works out the dependencies, network access and API keys a repo needs, then saves the environment for next time. It says keys are saved so that Claude never sees them. [The docs say](https://code.claude.com/docs/en/claude-code-on-the-web) these keys stay outside the sandbox and attach to matching requests after they leave the session.

**Where it comes from:** @ClaudeDevs, 2026-11-06.

## Claude Code 2.1.319

**What it is:** a Claude Code release from Anthropic.

**What it does:** when an MCP server runs a tool call as a task, Claude Code moves it to the background and tells Claude the result when it finishes. A new setting, `permissions.denyInsteadOfPrompting`, denies a tool call that would wait on a permission prompt, and the turn continues. That matters if you run agents with nobody watching.

**Where it comes from:** Anthropic. [The changelog](https://raw.githubusercontent.com/anthropics/claude-code/main/CHANGELOG.md).

## Stale bug comments cost coding agents time

**What it is:** a study from Google, by Sam Saccone and Madhuri Garikapati, of its own coding agent traces.

**What it does:** gives you a cheap thing to check. Agents met inline comments pointing at a bug ID in 13% of file-reading sessions and acted on the bug in 6% of those encounters, 83% of the time unprompted. Those detours added 70% more steps, 77% more tokens and 105% more wait time per session. 41% of them chased closed tickets.

**Where it comes from:** Google Developers blog, 2026-11-06. [The post](https://developers.googleblog.com/agent-distraction/).

## Cheap model first, Opus when unsure

**What it is:** a test by Hightouch on a 1,000-identity benchmark, sending a cheap model's uncertain cases to Claude Opus 5.5.

**What it does:** shows the pattern with numbers. On the easier test, Jev plus Opus scored 98.8% at $0.88 per 1,000, against 99.0% at $3.86 for Opus alone. On the harder test it scored 95.3% at $2.94, against 95.1% at $5.39. The band only helps if the cheap model knows when it is unsure: GPT-6 Luna's errors often fell outside it.

**Where it comes from:** Hightouch, 2026-11-04. [The post](https://hightouch.com/blog/llm-judges-identity-resolution).
