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

Tuesday, October 6UTC

  1. New
    OpenHandsOpenHands 1.25.0Agents

    OpenHands 1.25 adds model routing and automation integrations

    The OpenHands 1.25.0 release includes direct-prompt model-router settings, bulk model profiles and new automation template and native Git integration options. The official release also contains many fixes and UI changes; the listed features do not establish reliability across every backend.

    OpenHands1 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.

Monday, October 5UTC

  1. New
    CohereNorth 2Agents

    Cohere introduces North 2 with shared agents, skills and cost controls

    Cohere introduced North 2 as an enterprise AI platform with shared agents, reusable skills, memory, automations, model choice and administrative cost controls. It describes private deployment options including on premises and air-gapped environments. The announcement does not establish independent performance or universal self-serve access.

    Cohere1 source

  2. New
    QwenQwen Code 0.25.0Agents

    Qwen Code 0.25 adds local workspace-agent collaboration

    Qwen published Qwen Code 0.25.0 with local workspace-agent collaboration, A2A sharing, model broker controls and multiple hosted-run recovery features in its official release notes. The release artifact verifies the code changes; availability and reliability of managed hosted features still need hands-on testing.

    Qwen1 source

Saturday, October 3UTC

  1. New
    T3 CodeT3 Code nightly 2632Agents

    T3 Code nightly wakes agents when a pull request needs attention

    A server-side watcher checks linked pull requests once a minute and wakes their agent thread for check outcomes, new reviews or comments, and branch conflicts. Watch progress is persisted so an agent need not maintain its own polling loop. It stops on merge, closure or explicit removal and has limits for repeated comment wakes and host outages. A wake does not approve or merge the pull request.

    T3 Code2 sources

  2. New
    OpenAICodex 0.162.0-alpha.10Agents

    Codex alpha sends incremental tool catalogs and keeps MCP resource helpers available

    The alpha artifact incorporates Responses Lite catalog diffs and Code Mode resource-helper fixes. Changed tool declarations can be sent without repeating the entire catalog. These are prerelease runtime changes, not a claim that every account has access to Responses Lite or a new speed tier.

    OpenAI3 sources

  3. New
    DeepSeekDeepSeek Harness 0.2.1-alpha.1Agents

    DeepSeek Harness alpha tests Claude Code Mods compatibility and web task controls

    The prerelease introduces an experimental Mods compatibility layer and brings existing automation tasks into the web interface. It preserves unsent drafts and adds public-URL prefix support for reverse proxies. The invariant runtime plugin and its export path are removed, requiring extension updates.

    DeepSeek1 source

  4. New
    OpenClawOpenClaw 2026.9.8Agents

    OpenClaw 2026.9.8 adds GPT-6.1 Sol and changes how delegated results come back

    The stable release adds Sol to the OpenAI provider for accounts with access and returns delegated work to its original requester. Automatic peer reply loops are removed, so custom handoffs need explicit follow-ups. Windows upgrade repairs and lazy Codex catalog loading also ship.

    OpenClaw2 sources

  5. New
    T3 CodeT3 Code Orchestrator V2 nightlyAgents

    T3 Code V2 nightly connects cross-provider delegation, durable queues and scheduled tasks

    The first V2 nightly lets agents delegate to other providers, exposes thread controls through MCP, and supports Pi and OpenCode 2. Server-side queues and interrupted threads can survive restarts. Provider switching is lossy, and migration makes a one-way database copy; compatible mobile beta clients are required.

    T3 Code2 sources

  6. New
    MsgmateOpen-Chat GoAgents

    Open-Chat announces a self-hosted Go platform with MCP and background workers

    The author's announcement describes a local-first, multi-user platform with MCP integrations, asynchronous workers and sandbox orchestration. Current repository documentation confirms an AGPL core and separate proprietary components. Connecting cloud models or services still sends data outside the self-hosted runtime.

    Tim Schupp / Msgmate3 sources

Friday, October 2UTC

  1. New
    QwenQwen Code October 2 nightlyAgents

    Qwen Code nightly adds durable hooks, offline recovery and isolated runtime hosts

    Compared with the previous nightly, this prerelease adds durable Hosted hooks, offline workspace recovery bundles, immutable agent definitions and isolated remote Qwen runtimes. Code Mode gains freeform input and concurrent Bash calls. Earlier workspace collaboration is not a new launch.

    Qwen2 sources

  2. New
    ClaudeClaude Code 2.1.288Agents

    Claude Code 2.1.288 hardens permissions, session recovery and MCP execution

    The stable release repairs destructive shell-command approval and failed-hook checks, duplicate MCP calls, and context loss on resume. Plugin-defined team agents now keep their own tools and instructions; unattended stream recovery stops after three timeouts.

    Anthropic1 source

  3. New
    Builder.ioAgent-Native toolkit 0.200.0Agents

    Agent-Native 0.200 makes queued prompts and setup recovery durable

    The toolkit release preserves follow-up run options and keeps internal context out of visible chat. A turn refused because AI setup is missing retains its retry information server-side, with a one-tab resend claim that survives reloads. It also adds a stop hook and defaults direct Anthropic and OpenRouter connections to Sonnet 5.5. These are released library changes, not proof every hosted application has adopted them.

    Builder.io GitHub1 source

  4. New
    OpenClawOpenClaw 2026.8.35 extended-stableAgents

    OpenClaw extended-stable adds Sol support and repairs agent completion

    The gateway-only extended-stable release adds GPT-6.1 Sol compatibility and repairs delegated-answer delivery, suspended child capacity, and full cron-output recovery. Update handling preserves plugin configuration, while cron snapshot repair retains explicit tool allowlists. This is the maintained August branch, not the newest mainline release.

    OpenClaw GitHub1 source

  5. New
    Builder.ioAgent-Native desktop nightly 380Agents

    Agent-Native nightly connects durable chat queues with running-session recovery

    Builder.io desktop nightly 380 includes reviewed AgentKit changes that durably queue messages during an active run and turn run-slot conflicts into queued work. Built-in Agent-Native sessions recover active runs and reconcile saved messages after reload without duplicating streamed output. A separate CRM template change exposes seven authenticated, access-scoped reads to external agents, keeping direct writes off that surface. This is prerelease repository implementation, not proof every hosted app has rolled it out; multi-tab queue promotion remains unverified.

    Builder.io GitHub4 sources

  6. New
    OpenClawOpenClaw extended-stable 2026.8.34Agents

    OpenClaw extended-stable hardens upgrades without losing agent state or schedules

    OpenClaw 2026.8.34 is a gateway-only extended-stable maintenance release, not the newer latest branch. Its publisher lists 113 audit-selected fix units covering upgrade recovery, authentication, sessions, plugins and boundaries. The changes aim to preserve credentials, agent state, transcripts and schedules during repair, and tighten filesystem mutation, browser authentication, private skill ingress and runtime ownership. The matching npm package is published.

    OpenClaw GitHub2 sources

  7. New
    SupabaseSupabase ComputeAgents

    Supabase previews Compute for agent sandboxes and long-running services

    Compute is a private-alpha waitlist for agent execution environments and always-on services beside Postgres. It advertises API, MCP and CLI management with workload-scoped access. Preview terms limit it to internal evaluation, so this is not a generally available production hosting offer.

    Supabase2 sources

  8. New
    GrokGrok 4.7 on Ramp RouterAgents

    Grok 4.7 gets a 50% Ramp Router promotion through October 6

    Both SpaceXAI and Ramp Router announce the limited promotion. Router's public model catalog still displays baseline rates, so the announcement should not be converted into a guaranteed effective token price. This is a router-specific offer, not a permanent direct-Grok API price cut.

    SpaceXAI3 sources

Thursday, October 1UTC

  1. New
    EarendilPi 1.0Agents

    Pi 1.0 ships leaner Codemode, image-model calls and hardened MCP sign-in

    The MIT-licensed stable release adds image generation inside Codemode scripts and tightens MCP OAuth handling. The publisher reports a leaner tool prompt; this is not an independent cost benchmark. MCP itself existed before 1.0, while the terminal now defaults to fullscreen.

    Earendil2 sources

  2. New
    GrokGrok Primary BotAgents

    Grok Primary Bot can suggest work before you ask, without charging for the suggestion

    Grok Bot's October 1 announcement says the primary Bot can identify work it could take over and offer to help. The follow-up says suggestions do not count against usage. This does not establish that executing suggested tasks is free or that every account has received the rollout.

    Grok Bot / xAI2 sources

  3. New
    NotionNotion Agent connected ChatGPT planAgents

    Notion removes the matching-email restriction for ChatGPT plan sharing

    Notion's October 1 follow-up permits different email addresses when connecting a ChatGPT plan. Current help still requires Notion Business or Enterprise membership plus ChatGPT Plus or Pro. Funding applies only to Notion Agent sessions using a specific GPT model, not Auto, non-GPT models or Custom Agents. If ChatGPT usage runs out, switching to Notion credits requires a prompt.

    Notion3 sources

  4. New
    EarendilPi DurableAgents

    Pi Durable experiments with agents that resume after crashes

    Earendil released a separate experimental harness with persistent conversations, task checkpoints and concurrent forks. Interrupted tools rerun only when declared safe; other interruptions are reported to the model. Persistent storage is required for recovery, and the framework does not replace Pi's terminal coding agent.

    Earendil2 sources

  5. New
    ClaudeClaude Code at BarclaysAgents

    Barclays expands Claude and targets half its developers using Claude Code this year

    Anthropic announced an expanded Barclays rollout for software development, legacy modernization and operations. Barclays expects Claude Code adoption to reach half its developer population by the end of 2026 and a majority in 2027. The announcement describes an existing knowledge assistant and email-routing deployment, with governance and human oversight. These are publisher-reported deployments and targets, not independently verified results or new Claude capabilities.

    Anthropic1 source

  6. New
    PerplexityPerplexity Computer Business SkillsAgents

    Perplexity adds ten business Skills for eligible American Express users

    Perplexity announced reusable Computer Skills for finance, marketing, operations and hiring. Access requires an eligible U.S. American Express Business Card, an active auto-renewing Perplexity Enterprise subscription and card linkage through Plaid. Skills collect task details, request missing connections and produce outputs for owner review. They consume Computer Credits; a full cash-flow forecast also needs the operating bank account connected. This is not a free or worldwide benefit.

    Perplexity1 source

  7. New
    Strands AgentsStrands Decider 2BAgents

    Strands releases a local 2B decision model for agent routing and guardrails

    Strands Decider chooses among supplied options and returns confidence scores instead of generating free-form text. The Apache-2.0 code and model checkpoint support local CPU, CUDA and MPS experiments. The model card warns that calibration is domain-limited and long, multi-step questions remain a weak spot.

    Strands Agents4 sources

Wednesday, September 30UTC

  1. New
    RunwayPraxis-1Agents

    Runway previews Praxis-1, turning video pretraining into robot actions

    Praxis-1 is being tested on partner hardware before public release. Runway plans open weights later; they are not available merely because the announcement calls it an open-weight model.

    Runway1 source

  2. New
    GoogleGemini SkillsAgents

    Gemini chat gets stackable skills and a migration path from Gems

    Google is rolling reusable skills into Gemini chat with slash invocation, reference files and multiple skills combined in one task. Workspace rollout and sharing arrive later. Gems will be retired on different schedules for personal and Workspace accounts, with automatic migration; Google says the separate Labs mini-app experiment Opal will also wind down in November.

    Google1 source

  3. New
    OpenClawOpenClaw 2026.9.7Agents

    OpenClaw 2026.9.7 adds OpenAI Agents API and rollback-safe agent updates

    OpenClaw 2026.9.7 adds an Agents API plugin for OpenAI-hosted or self-hosted environments with streaming, steering, tools, attachments and preserved tool history. The same release backs up state and agent databases before migrations, restores them on rollback, and improves in-flight turn continuity across Gateway restarts.

    OpenClaw GitHub2 sources

  4. New
    IBMwatsonx Orchestrate runtime evaluationAgents

    IBM makes per-evaluator sampling controls available for live agent traffic

    IBM identifies September 30 as general availability for runtime LLM-as-a-Judge controls. Six built-in evaluators can be enabled independently, with a 3% default trace sample and per-evaluator settings up to 20%. The October 2 recap also describes older identity and multi-cloud previews; those are not newly generally available features.

    IBM2 sources

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