Skip to main content

The news desk

AI news, checked against the source.

What changed in AI, one story at a time. Each one links to its primary source, and facts stay apart from interpretation.

Last updated

Monday, October 5UTC

  1. New
    GitHubReviewBenchDeveloper

    GitHub releases ReviewBench to evaluate AI code review agents

    GitHub launched ReviewBench as a research preview for evaluating AI code review agents against a public benchmark corpus and structured findings. GitHub says its benchmark reflects representative pull requests and offers a self-serve runner, methodology and grader configuration. Its validation and production-correlation claims are GitHub assessments.

    GitHub1 source

Free, email only

Unlock the full brief free.

  • Thirty days of stories to browse, not seven
  • On every story: what to check, each source and why it matters
  • The same signup unlocks every free tool

This is not an account: there is no password, and the unlock is a cookie in this browser. You also join the Rise Productive newsletter from Demetri Panici, about once a week: what I built and what changed in AI. We'll email you a link to confirm, and you can unsubscribe in one click. The same signup unlocks every free tool on the site. How your email is handled.

Already subscribed? Enter the same email to unlock this browser. You won't be signed up twice.

Sunday, October 4UTC

  1. New
    T3 CodeT3 Code nightly 2644Developer

    T3 Code nightly preserves delegated work across restarts

    The October 4 T3 Code nightly improves preservation of delegated tasks and queued threads across restarts, cleans up stale development servers when providers change, and adds recovery for archived native Codex sessions and workspace preparation. Its mobile Working section is still a beta. These are prerelease workflow fixes, not a new underlying model or a guarantee of failure-free unattended work.

    T3 Code1 source

Saturday, October 3UTC

  1. New
    ClaudeClaude Code 2.1.289Developer

    Claude Code 2.1.289 tightens permission rules and plugin boundaries

    This stable Claude Code update fixes permission handling for compound commands, symlink reads and shell environment prefixes. It also stops plugins from replacing organization MCP sign-in descriptions and adds agent spawning/state APIs. These changes matter to teams relying on approval and deny rules, but do not establish that every security risk is resolved.

    Anthropic1 source

  2. New
    QwenQwen Code October 3 nightlyDeveloper

    Qwen Code nightly adds email channels and trusted reboot recovery

    The October 3 Qwen Code nightly includes an IMAP/SMTP email channel, durable local process support with trusted reboot recovery, and more tightly scoped microphone consent. These changes are verified against the previous nightly rather than treating its cumulative release notes as entirely new. This is a prerelease coding-agent update, not a new Qwen model.

    Qwen1 source

  3. New
    T3 CodeT3 Code nightly 2638Developer

    T3 Code nightly fixes mobile threads and stuck Claude plan sessions

    This T3 Code prerelease improves mobile thread scrolling and subagent details, repairs stuck Claude plan sessions and adds Mermaid diagram rendering. It also adjusts usage scanning. The update is a workspace reliability release, not a new underlying model or guaranteed speed increase for every project.

    T3 Code1 source

  4. New
    MiniMaxOpenAgentCore 0.0.9Developer

    OpenAgentCore 0.0.9 makes Compose startup self-contained

    The release bundles initialization metadata with its pinned init image and publishes Web from one Compose file. This removes a startup download dependency and the separate ports overlay. It also repairs generated admin-key handling; public TLS and authentication remain operator responsibilities.

    MiniMax3 sources

  5. New
    MiniMaxOpenAgentCore 0.0.8Developer

    OpenAgentCore 0.0.8 moves public API ingress into Web and hands HTTPS to the operator

    The 0.0.6-0.0.8 changes remove the managed domain/gateway service and route API calls through Web. Public deployments now rely on the hosting platform or an operator-managed reverse proxy for TLS. The update also permits explicit HTTP LAN addresses; that is not a safe default for public exposure.

    MiniMax4 sources

  6. New
    GoogleGemini CLI October 3 nightlyDeveloper

    Gemini CLI nightly fixes Enter and Space confirmation in selection lists

    The October 3 prerelease repairs selection-list confirmation. This is a focused terminal interaction fix, not a model launch or a new autonomous capability.

    Google1 source

Friday, October 2UTC

  1. New
    MiniMaxOpenAgentCore 0.0.5Developer

    OpenAgentCore 0.0.5 packages self-hosted agent infrastructure for Compose

    The release includes self-contained Compose templates for the control plane, persistent storage, credential initialization and ordered database migrations. Startup can use localhost before a public origin is assigned. The v0.0.5 template still pins its container images and installer payload to v0.0.3; this is packaging work, not a new model or runtime capability.

    MiniMax-AI4 sources

  2. New
    n8nn8n 2.41.6Developer

    n8n 2.41.6 keeps a rejected promise from taking down its task runner

    The stable patch adds an unhandled-rejection listener to the JavaScript task runner. Instead of ending the runner and failing other in-flight tasks, it records a bounded warning and keeps the process alive. This is distinct from the previously covered beta fix for queue-completion recovery.

    n8n2 sources

  3. New
    OpenAICodex 0.162.0-alpha.4Developer

    Codex alpha adds managed worktree tools without moving the current task

    Codex 0.162.0-alpha.4 includes feature-gated worktree creation, status and listing tools for trusted local projects. Creation is asynchronous; task attachments persist across restarts, while the current working directory and uncommitted source changes stay in place. The prerelease also adds sandbox-checked file-write streaming. Its native cloud attach client has no automatic reconnect, and an optional Guardian comparison does not change approvals. These are experimental changes, not universally enabled stable features.

    OpenAI GitHub5 sources

  4. New
    GoogleGemini CLI October 2 nightlyDeveloper

    Gemini CLI nightly makes long sessions lighter and restores reliable interruption

    The October 2 Gemini CLI nightly includes append-only history deltas and a rolling in-memory window, reconstructing older messages from its backing log rather than deleting them. Extension state gains atomic writes and backup recovery. Ctrl+C can again cancel an active request even with text in the input buffer, and ACP session loading resolves exact IDs while rejecting traversal. The fixes are verified in a prerelease artifact, not a guarantee that external tool side effects can be undone.

    Google GitHub5 sources

  5. New
    OpenAIGPT-6 production guideDeveloper

    OpenAI publishes a practical GPT-6 guide for production agents and long-running work

    The guide connects model selection with task-level evaluations, caching, compaction, reasoning effort, and explicit instructions about completion and approval. It also explains steering, asynchronous tools, and delegation for long-running work. This is implementation guidance for existing GPT-6 models, not another model launch.

    OpenAI1 source

  6. New
    SupabaseAgent-ready local developmentDeveloper

    Supabase tests Docker-free local stacks for coding agents

    The announced native local stack is alpha and off by default, with separate instances per directory. Declarative Schemas 2.0 uses pg-delta to derive migrations from schema files, while config pull brings dashboard settings into the repository.

    Supabase3 sources

Thursday, October 1UTC

  1. New
    OpenAIOpenAI Python SDK 3.23.0Developer

    OpenAI Python SDK adds typed agent answers, file staging and turn artifacts

    The official SDK release adds session-stream answer helpers, file staging and artifact downloads. It also exposes typed application actions as beta agent tools. These client-library additions do not grant organization access to the underlying APIs or remove approval requirements.

    OpenAI1 source

  2. New
    OpenAICodex 0.162.0-alpha.1Developer

    Codex alpha adds gated subagent tool inheritance and keeps user restrictions in handoffs

    Codex 0.162.0-alpha.1 contains a disabled-by-default flag that lets fresh V2 subagents inherit client-defined dynamic tools without copying parent history. Its reviewed changes also retain root user restrictions in Guardian handoff context, subject to existing message limits, and follow MCP tool-list pagination in legacy protocol mode. These are verified prerelease implementation changes, not universally enabled capabilities or a new model.

    OpenAI GitHub4 sources

  3. New
    QwenQwen Code October 1 nightlyDeveloper

    Qwen Code nightly adds Hosted approval controls, with important visibility limits

    The October 1 Qwen Code nightly includes WebShell allow-and-deny cards for Hosted sessions that advertise Actions. Only the session creator may answer. The current Hosted producer does not supply tool arguments, so the card explicitly reports them unavailable. Separate gated work exposes durable shell results through authenticated paging and streaming, while scheduled prompts now persist in transcripts. This does not enable the unfinished Managed engine for all users.

    Qwen GitHub5 sources

  4. New
    OpenAICodex 0.160.0Developer

    Stable Codex adds opt-in Guardian history and safer session recovery

    Codex 0.160.0 ships opt-in Guardian retrieval of earlier user instructions and handoff-aware review context. Both features are disabled by default and remain bounded by current parent policy. The stable release also restores saved session permissions, supports projectless sessions when policy permits, and resumes unsent queued messages only after uncertain submissions are resolved. This is a stable release, distinct from the newer experimental alpha line.

    OpenAI GitHub3 sources

  5. New
    ClaudeClaude Code ModsDeveloper

    Claude Code Mods let plugins reshape the agent, with an opt-in side agent to catch misses

    Claude Code 2.1.287 adds Mods: JavaScript or TypeScript handlers that can add interface panes and commands or intercept session events. The release also adds You should know, an opt-in side agent for first-party sessions with telemetry enabled. Mods run with user privileges; the docs explicitly warn that they can reach secrets and approve tool calls. This is deeper customization, not a sandboxed prompt pack.

    Anthropic GitHub4 sources

  6. New
    OpenAICodex 0.161.0-alpha.13Developer

    Codex alpha renews enterprise MCP credentials without replaying refused operations

    A merged change in Codex alpha.13 adds renewable enterprise MCP HTTP authentication. It confines bearer tokens to the configured endpoint, rejects redirects, serializes renewals and supports revocation. A 401 becomes an authentication-required error instead of automatically replaying the operation; delayed failures preserve newer credentials. The artifact also repairs asynchronous Guardian context ordering. These are prerelease reliability changes, not general access or a new model.

    OpenAI GitHub3 sources

  7. New
    OpenAICodex 0.161.0-alpha.11Developer

    Codex alpha adds per-turn Cyber access selection and richer delegation reviews

    The October 1 Codex alpha artifact includes an experimental per-turn Cyber access-program option for exec and the TypeScript SDK. Merged changes validate program access and filter it for the built-in OpenAI provider. Guardian sender reviews gain preceding assistant context, explicitly treated as untrusted context rather than authorization. Temporary structured threads also inherit the server's approval policy.

    openai/codex4 sources

  8. New
    n8nn8n 2.42.2 betaDeveloper

    n8n beta fixes workflow runs stuck after missed queue completion events

    n8n 2.42.2 replaces the main process's Bull job.finished() wait with a JobOutcomeTracker. Worker messages, Bull completion or failure events, and a once-per-minute database status check can settle a wait. The repair targets leaked listeners, timers, active executions and webhook requests after a missed Redis completion event. It is a beta release, not a new agent product or a promise of exactly-once execution.

    n8n-io/n8n2 sources

  9. New
    GrokGrok for IntuneDeveloper

    Grok adds a separate iOS app for Intune-managed workplaces

    The X Corp. App Store listing describes Grok for Intune as an organization-managed edition with work-email sign-in and support for Microsoft Intune app-protection policies. IT administrators control deployment and restrictions on saving, sharing and data transfer. The listing also includes existing Grok research, Imagine and voice capabilities, whose availability may vary by organizational policy. This is a managed-app distribution release, not a new generation model.

    X Corp. on App Store2 sources

  10. New
    TencentAI-Infra-Guard 4.6.4Developer

    AI-Infra-Guard adds a composable prompt-injection research toolkit

    Tencent's 4.6.4 release vendors pikit into its research tools and forwards optional reasoning effort through Skill-Scan. Pikit documents controlled attack, channel and defense combinations and integrations for agent frameworks, OpenClaw and Hermes. This is an authorized research tool, not evidence that a particular deployed agent is vulnerable or protected.

    Tencent/AI-Infra-Guard2 sources

  11. New
    PostHogPostHog Desktop 0.61.621Developer

    PostHog Desktop separates OAuth tokens from agent CLI environments

    The new desktop release includes a fix that gives agent CLIs placeholder credentials while a local auth proxy supplies the real OAuth credential. It also adds selected-block context and scoped canvas comments, a Sonnet 5.5 catalog option, and gateway routing behind feature flags. These are PostHog integration changes, not an Anthropic model announcement.

    PostHog/posthog2 sources

  12. New
    Builder.ioAgent-Native Core 0.198.8Developer

    Agent-Native 0.198.8 restores app actions for development subagents

    Core 0.198.8 exposes application actions to dev-mode subagents and keeps optional migration codemods out of application build graphs. The linked implementation identifies the app-action change as a restoration. Companion package version bumps do not establish separate product launches.

    BuilderIO/agent-native2 sources

  13. New
    GoogleGemini CLI 0.64 nightlyDeveloper

    Gemini CLI nightly protects parallel agent edits with atomic file writes

    The October 1 prerelease serializes file-tool operations and writes files atomically to address lost updates and shadow Git collisions when tools or subagents work concurrently. It also fixes a headless CPU hang caused by scoped-package references and quoted code. The changes are in a nightly artifact, not a new stable release.

    google-gemini/gemini-cli3 sources

  14. New
    DigitalOceanAgent Development KitDeveloper

    DigitalOcean deprecates ADK and stops accepting new ADK agent deployments

    The gradient-adk Python package and gradient CLI are deprecated for new agent deployments. Existing ADK agents continue to run and incur model-usage charges. DigitalOcean has not yet published its migration guidance or end-of-support date.

    DigitalOcean2 sources

  15. New
    JetBrainsAir in JetBrains IDEs EAPDeveloper

    JetBrains opens Air in IDEs EAP for bringing existing coding agents into one workspace

    JetBrains opened the Air in IDEs Early Access Program through a Marketplace plugin and 2026.3 EAP builds. Air organizes existing agents and subscriptions rather than supplying an AI provider: users install and configure their agents. The tool window coordinates sessions, costs and worktrees, and supported agents can use IDE capabilities such as debugging and semantic search. Cloud runs require JetBrains AI access, while mobile support is still coming. This is a concrete EAP availability change after the earlier Air introduction, not another copy of that event.

    JetBrains1 source

  16. New
    DigitalOceanInference and Agents balanceDeveloper

    DigitalOcean adds a separate prepaid balance for inference and agents

    Serverless Inference and Managed Agents can draw from a dedicated non-expiring balance, funded in $5-$500 increments. Other products cannot spend it. Optional auto-reload creates recurring charges; eligible monthly agent plans are separately offered in public preview.

    DigitalOcean3 sources

Days are UTC. Some sources report an exact time and others only a date; both are placed on their UTC day.