AI News Nov 19, 2026 14 min read

Claude Code Code Mode runs JS tool scripts up to 2 hours

Claude Code Code Mode runs JavaScript tool scripts in v2.1.323+, alongside Rumble-2-Flash speech, AI Studio Android, Codex CLI 0.170.0, and Google SI…

Thu, Nov 19, 2026Issue #25 stories15 min read

AI developer tools added faster script execution, voice interfaces, and mobile app backends this week, while researchers tested ways to carry agent optimization logs across models. Official docs and release notes establish the new flags and features, while speech accuracy numbers and cross-model agent gains still rely on company benchmarks and an unreviewed preprint.

Get the weekly AI briefing by email. Free.

Coding Agents and Experimental Modes01 / 05

Claude Code Code Mode lets Claude bundle tool calls in JavaScript

Layered paper terminal and script scroll representing Claude Code Code Mode

As of November 18, Anthropic’s Claude Code documentation details Code Mode, an experimental feature where Claude writes one JavaScript script to run multiple tool calls in a single step.

Read the full story~3 min

Normally, Claude calls tools like Bash, Read, or Edit one by one and reads each output before choosing the next action. Code Mode lets Claude batch those steps into one script, which matters for multi-file edits, filtered searches, and parallel commands.

  • Activation and version: Requires Claude Code v2.1.323 or later (native build) with the environment variable CLAUDE_CODE_EXPERIMENTAL_CODE_MODE set to 1.
  • Model and surface limits: Supports Opus 5.5, Sonnet 5.5, Haiku 5.5, and Fable 5.2 in interactive terminal sessions only, excluding Desktop, VS Code, and cloud gateways.
  • Execution timeout: Waits up to 10 minutes by default, or runs in the background for up to 2 hours when started in auto or bypassPermissions mode.

What changes

Terminal users on direct Anthropic accounts can run long builds or multi-file edits in fewer round trips while keeping the conversation responsive in auto mode.

Limits

Code Mode is experimental, off by default, and may be removed in any release. The documentation page lists no publication date and excludes Bedrock, Google Cloud, Microsoft Foundry, HIPAA, and zero data retention setups.

Community reaction

■A Threads user highlighted that long test suites can run in the background for up to two hours without blocking the chat when Code Mode starts in auto or bypassPermissions mode, subject to mode and launch-flag conditions.

■The same thread walked through the four-line retry example from the docs, noting that Ctrl+O is needed to inspect the hidden JavaScript.

■The thread’s author emphasized the v2.1.323 minimum version and warned that the feature does not work in the Desktop app or VS Code panel.

Big picture: Developers running heavy terminal workflows can test batched tool execution today, provided they stay on direct Anthropic accounts without third-party gateways.

How batched scripts change tool execution and session visibility

According to the Code Mode guide, a single script can search files and run tests before returning only the final output Claude needs. In the example shown, Claude finds a retry policy file, raises maxAttempts from 3 to 5, and runs tests in four lines of JavaScript.

The conversation view still displays standard diffs and tool results rather than raw code. Users who want to audit the generated JavaScript script must press Ctrl+O to open the transcript viewer and expand the collapsed Code row. Sessions started with Code Mode can only be resumed with the flag turned on, as explained in the Code Mode guide.

Permission checks, billing, and environment configuration

Each tool call inside a script still passes through deny rules, ask rules, and PreToolUse hooks, which can inspect the script via the REPL tool’s code input. In auto mode, a classifier reviews the script before execution rather than each individual tool call, and it cannot see values computed at runtime, per the Tools reference.

As detailed in the Auto mode classifier billing docs, server-side classifier checks run at no extra charge wherever enabled, though availability depends on platform, region, credential, and rollout even without a proxy. When server-side checks do not reach a session, such as behind an incompatible proxy, Claude Code falls back to client-side classifier requests that count toward token usage on accounts receiving the fallback notice, while Pro, Max, Ultra, and Team plans never see that notice. Teams can set or block the toggle using the env key in settings files, following the Environment variables reference.

Version requirements and compatibility boundaries

Code Mode requires Claude Code v2.1.323 or later for single-session terminal use. Running it alongside agent teams requires v2.1.326 or later in the default in-process display mode. Earlier coverage in our previous briefing noted that v2.1.326 added agent team support and tool option warnings.

Users cannot move a Code Mode session to the Desktop app, use Remote Control, or run computer use and Claude in Chrome. Organizations can disable the feature company-wide by setting CLAUDE_CODE_EXPERIMENTAL_CODE_MODE to 0 in managed settings, though admins cannot force it on for users, according to the Code Mode documentation.

Read more · Run tool calls from code with Code Mode ↗ · Environment variables ↗ · Auto mode classifier request charges ↗ · Tools reference ↗

Open Speech and Mobile App Builders02 / 05

Rumble-2-Flash brings 0.5B open speech recognition to 40+ languages

Paper-cut microphone and stopwatch representing Rumble-2-Flash speech recognition
6.96% average WER7.86% WER on phone calls

On November 18, Reson8 published Rumble-2-Flash on Hugging Face, releasing a 0.5-billion-parameter automatic speech recognition (ASR, speech-to-text) model under the Apache 2.0 license.

Read the full story~2 min

The model uses the Qwen3-ASR architecture so teams can run it natively in Hugging Face Transformers or vLLM without custom code. It targets developers who want fast, permissive multilingual transcription without deploying a large model.

  • Languages and controls: Supports 40+ languages with automatic language detection, plus formatting controls for digits, fillers, verbatim repeats, and forced language.
  • Distillation method: Trained from Reson8’s larger Rumble-2 model using on-policy distillation with GOLD (a method that aligns differing teacher and student tokenizers).
  • Reported accuracy: Reson8 reports 6.96% multilingual WER (word error rate) on the Open ASR Leaderboard and 7.86% WER on a 302-hour call-center benchmark.

What changes

Developers can download Apache 2.0 weights from Hugging Face to run on standard Transformers and vLLM stacks from day one, or use the model through Reson8’s API.

Limits

All accuracy and speed metrics come from Reson8’s announcement and were not independently verified here. The blog excerpt does not list all 40+ supported languages, leaving Korean support unconfirmed.

Community reaction

■No separate X or Threads posts about Rumble-2-Flash appeared in the reviewed social reaction collection.

■The Hugging Face announcement post showed 3 upvotes shortly after publication.

■Reson8 asked developers to share projects and version requests in the model’s Discussions tab.

Big picture: Small open-weight speech models are closing the gap with 30B-class systems on noisy audio, though teams should benchmark their own target languages first.

How on-policy distillation with GOLD bridges different tokenizers

In its release post, Reson8 explains that Rumble-2-Flash was distilled from its larger production model, Rumble-2, rather than trained solely on static transcripts. During on-policy distillation, the student transcribes audio itself while the teacher supplies next-token probabilities at each position for the student to match, providing token-level feedback on the mistakes the student makes.

Because adopting Qwen3-ASR meant using a different vocabulary than Rumble-2, Reson8 applied GOLD to align tokenizers across differing text splits. The AnySpeechRecognizer announcement notes that Qwen3-ASR covers 30 languages and 22 Chinese dialects in 0.6B and 1.7B sizes. However, it typically runs slower on GPU for file transcription than smaller Core ML engines.

Benchmark claims across clean speech and noisy phone calls

On the Open ASR Leaderboard multilingual track, Reson8 reports a 6.96% average WER across nine languages (Chinese, German, French, Italian, Spanish, Portuguese, Hindi, Dutch, and English). The company says that beats Meta’s omniASR-LLM 7B and Whisper large-v3. On the English track, it reports a 6.6% average WER. Separately, Reson8 reports throughput above 7,000 RTFx (real-time factor), transcribing an hour of audio in about half a second, when batches stay full on the larger datasets in Hugging Face’s Open ASR Leaderboard.

To test compressed audio, Reson8 evaluated AppTek’s 302-hour call-center dataset across 14 English accents using AppTek’s scoring script. The company reports 7.86% WER on phone calls, nearly matching NVIDIA’s 30B Nemotron 3.5 Transcribe at 7.78% and beating Qwen3-ASR 0.6B at 8.7%, while trailing Qwen3-ASR 1.7B at 7.4%, per the Reson8 blog.

Read more · Rumble-2-Flash: an open 0.5B speech model for 40+ languages ↗ · Introducing AnySpeechRecognizer and AnyDecisionModel ↗

Open Speech and Mobile App Builders03 / 05

Google AI Studio Build adds full-stack Android apps and custom domains

Paper-cut smartphone on a cloud pedestal representing Google AI Studio Android apps

On November 17, Google announced on the Google Developers Blog that Google AI Studio Build now generates full-stack Kotlin Android apps with Firebase backends, an upgraded browser emulator, and Google Play publishing.

Read the full story~3 min

Powered by Gemini 4 Carbon and Hydrogen alongside the Antigravity Agent and the official Android CLI, Build mode turns natural-language prompts into multi-screen Android apps. Separately, the AI Studio Build documentation notes that users can buy and connect custom domains directly when publishing on a paid Google Cloud project.

  • Firebase backend setup: Automatically configures Cloud Firestore with Security Rules, Google Sign-In, Firebase AI Logic, and Firebase App Check so API keys stay off the APK.
  • Emulator and UI tools: Adds faster startup, 1200p HD streaming, live audio output, custom fonts, and one-click emulator screenshots sent directly to the coding agent.
  • Play and domain publishing: Lets users publish signed .aab bundles to a personal Google Play library or internal testing track, while web apps on paid Cloud projects support custom domains.

What changes

Developers and solo builders can prototype, test with audio and 1200p video in a browser, and ship signed Android builds to their phones without a local Android Studio setup.

Limits

The blog post does not mention custom domains, so when that feature launched is not disclosed. Official docs reviewed here do not confirm claims that Android downloads are restricted to the US, and Android builds remain client-side only without server-side runtimes.

Community reaction

■Threads users noted that Google’s ecosystem now spans models, coding, emulator testing, and Play distribution, though some said frequent updates are hard to track.

■Social posts on X and Threads claimed Android device downloads and personal Play publishing are limited to US users for now.

■X commenters reacted to custom domain purchases in Build mode by wishing for .test domains, while others noted RFC 2606 reserves .test for local testing.

Big picture: Prompt-to-app tools are moving past static prototypes by wiring authentication, databases, and app-store signing into one browser workflow.

How Firebase provisioning and the Android emulator work together

According to the Google Developers Blog, the Build agent detects when a prompt needs persistence, sign-in, or AI calls. With user confirmation, it provisions Cloud Firestore, Firebase Authentication, and Firebase App Check protection so Gemini API keys stay off client APKs, using Google Fraud Defense (reCAPTCHA Enterprise) for Play release builds, while the Android guide notes that emulator, USB, and tester debug builds use an App Check debug token.

Inside the browser, the upgraded emulator boots faster and adds a 1200p HD video toggle alongside real-time audio output. Developers can click Take a screenshot to attach a live frame directly to their prompt for UI fixes. Under the hood, the official Android CLI handles project scaffolding and Gradle builds.

Publishing paths, emulator limits, and billing tiers

The Build Android apps guide notes that Android apps in AI Studio remain client-side only. Features requiring a server-side runtime, such as Google Workspace APIs and multiplayer, work only in web apps. The cloud emulator also excludes camera, NFC, Bluetooth, real GPS, and Google Play services, though developers can install APKs over USB via WebUSB without local ADB.

For distribution, teams can publish to the Google Play Console internal testing track for up to 100 testers with a one-time $25 developer fee, per the Android guide. On the web side, the Build mode docs state users can buy and connect custom domains on paid Google Cloud projects.

According to the AI Studio billing docs, eligible accounts can use the Google Cloud Starter Tier to publish up to 2 full-stack apps without a billing account. Moving to Paid Tiers requires linking a billing account; accounts on the Prepay plan (which new users default to) must prepay at least $5 in credits that expire after 12 months, whereas some accounts are offered or temporarily assigned a Postpay plan.

Read more · Build full-stack Android apps with Gemini in Google AI Studio ↗ · Build apps in Google AI Studio ↗ · Build Android apps in Google AI Studio ↗ · Billing ↗

Terminal Clients and Agent Memory Research04 / 05

Codex CLI 0.170.0 adds TUI voice calls and MCP auth fixes

Mechanical keyboard and audio wave badge representing Codex CLI 0.170.0

On November 18, OpenAI published Codex CLI 0.170.0 on GitHub (tagged `rust-v0.170.0`), introducing WebRTC voice call negotiation in the terminal user interface (TUI), refreshable telemetry headers, and experimental multi-environment MCP support.

Read the full story~2 min

The release focuses on terminal responsiveness, observability, and connection stability for developers running Codex CLI locally or against cloud sessions. It also resolves several sandbox permission bugs on Windows and Linux.

  • TUI voice and telemetry: Adds WebRTC negotiation for realtime voice calls in the TUI and automatically refreshed authentication headers for OpenTelemetry HTTP exporters.
  • Same-name MCP servers: Adds experimental support for keeping identically named MCP (Model Context Protocol, a standard way for AI tools to connect to external apps) servers separate across environments.
  • Reliability and sandbox fixes: Fixes MCP OAuth credential races, oversized WebSocket HTTP fallbacks, Cloud conversation reconnection, and Windows sandbox permission preservation.

What changes

Codex CLI users get groundwork for realtime voice calls in the terminal, cleaner MCP OAuth handling, and fewer dropped Cloud sessions after transient stream loss.

Limits

The release page itself shows no visible date; the November 18 date comes from GitHub’s release record. It does not explain how to enable TUI voice calls, which pricing tiers support them, or when the experimental MCP feature will stabilize.

Community reaction

■No separate X or Threads posts about Codex CLI 0.170.0 appeared in the reviewed social reaction collection.

■On the GitHub release page, 8 users left emoji reactions, with thumbs-up and rocket receiving 4 each.

■The changelog highlights contributions from 22 contributors, including euroelessar and GrantBirki.

Big picture: Terminal coding agents are adding voice and better enterprise telemetry while tightening the OAuth and sandbox plumbing underneath.

Realtime voice negotiation, telemetry headers, and MCP separation

According to the Codex CLI 0.170.0 release notes, three new features headline the `rust-v0.170.0` release. First, pull request #61172 adds WebRTC voice call negotiation inside the TUI, laying the protocol groundwork for realtime voice sessions in the terminal. Second, pull request #61028 introduces automatically refreshed authentication headers for OpenTelemetry HTTP exporters so long-running sessions keep exporting traces.

Third, pull request #60905 adds experimental support for identically named MCP servers across different environments, keeping their configurations separate. The release page does not provide user-facing flags, pricing details, or setup instructions for TUI voice calls, leaving rollout scope and availability unconfirmed beyond the code merge.

WebSocket fallbacks, OAuth races, and Windows sandbox fixes

The changelog for 0.170.0 also bundles stability and security fixes across transport and sandboxing layers. Oversized WebSocket requests now fall back to HTTP immediately without unnecessary retries (#60840, #60841), and idle Cloud conversations reconnect automatically while preserving active recovery turns (#60969, #61014).

For tool authentication, OpenAI fixed MCP OAuth credential races during cancellation, login, and logout (#60849, #60981). On the filesystem side, the release preserves file permissions during repeated Windows sandbox setups (#60943), stops child threads from losing existing sandbox restrictions (#61134), and updates sandbox documentation around permission profiles (#61133). The changelog also notes Guardian context budget reservations for Code Mode results (#60914).

Read more · Release 0.170.0 · openai/codex ↗

Terminal Clients and Agent Memory Research05 / 05

Google preprint tests Self-Improvement Memory for agent harnesses

Bound notebook with tabbed dividers representing Self-Improvement Memory

On November 13, researchers from Google Cloud AI, Google DeepMind, and Penn State submitted an arXiv preprint proposing Self-Improvement Memory (SIM), a structured log that helps AI optimizers keep rewriting agent harnesses across new models and benchmarks.

Read the full story~2 min

An agent harness manages the context, tools, and execution loop around a model. Automated harness tuning is often studied within a fixed model and benchmark setting, and when either changes, code alone may not explain why past edits were made or how they performed.

  • How SIM works: Groups harness edits by recurring root-cause hypotheses and records each attempt’s rationale, predicted and observed outcomes, and provenance.
  • Coding benchmark gains: Across 10 independent source runs moving from Terminal-Bench 2.1 to Terminal-Bench 4, SIM improved 9 of 10 paired continuations by an average of 2.36 points (p = 0.003).
  • Transfer caveats: An Opus-evolved workspace harness started at 68.41 on Gemini, below the seed harness score of 73.89, showing that inherited edits can initially hurt a new model.

What changes

Teams researching automated agent optimization can pair evolved harnesses with structured edit histories rather than discarding tuning progress after a model upgrade.

Limits

This is an unreviewed arXiv preprint, not a peer-reviewed paper or commercial product. Only the abstract through Section 4.1 was reviewed here, and the excerpt refers to Opus and Gemini without confirming exact version numbers.

Community reaction

■X posts called the preprint a sign that recursive self-improvement is getting closer because new models no longer have to start harness tuning from zero.

■Commenters highlighted the 9-out-of-10 coding win rate and cited specific model names and workplace scores not shown in the reviewed preprint excerpt.

■Other replies praised the researchers’ work on the paper.

Big picture: Carrying the reasoning behind past harness changes could make model upgrades less disruptive, though the approach remains early research.

Why harness code alone fails when models or benchmarks change

In their November 13 preprint, the authors define Continual Recursive Self-Improvement (CRSI) as continuing automated harness optimization after a benchmark or policy model changes. When a model is swapped, an evolved harness loses context because the code itself does not show why an edit was proposed or which failures it solved.

Without that history, an optimizer may repeat failed experiments or strip away useful safeguards. For example, in a workspace experiment reported in the arXiv paper, an Opus-evolved harness started at 68.41 on Gemini, below the 73.89 score of the unedited seed harness. Similarly, in text classification, a source-evolved FewShotAll harness started at 72.92 on the target suite, below the seed’s 75.72.

How Self-Improvement Memory tracks root causes and outcomes

To preserve that tuning history, SIM stores edits under recurring root-cause hypotheses rather than letting the task-solving policy rewrite itself directly. Each entry logs the modification rationale and outcomes, including predicted fixes, observed task results, and provenance across domains, benchmarks, policies, and rounds, per the preprint methodology.

Across four domains, Full SIM outperformed harness-only and root-cause-only continuation in six cross-dataset settings and four Opus-to-Gemini settings. In coding (Terminal-Bench 2.1 to Terminal-Bench 4), paired continuations with SIM improved 9 of 10 source runs by an average of 2.36 points (p = 0.003). The paper is posted under a CC BY 4.0 license, which per arXiv’s license guide permits commercial reuse and adaptation with attribution.

Read more · Self-Improvement Memory Enables Continual Recursive Self-Improvement ↗ · Licenses – arXiv info ↗

Fluxaivory desk

Built from primary sources with AI assistance and published after automated checks. AI Use Policy · Corrections

Follow Fluxaivory

Get the briefing by email

A weekly roundup of the top AI news in plain English, from primary sources. Free. Unsubscribe anytime.