This week in AI tools is not just another pile of demos. Andrew Warner and Corey Ganim at The Next New Thing walked through 16 launches that point in the same direction: AI is moving out of the chat box and into the places work already happens.
NotebookLM is turning sources into short vertical videos. xAI is making phone-call agents easier to configure. Marketing agents are moving into Slack. Agent harnesses are becoming open source. Coding agents are going mobile. Hermes is making context usage visible. That is the real pattern.
Source Note
Credit for the source roundup goes to The Next New Thing, Andrew Warner, and Corey Ganim. The episode link is Free Claude Tag + agents battle + TikTok video maker + more.
The 16-link resource list comes from The Next New Thing's Links We Covered PDF for July 2, 2026. Where available, I also checked official product pages and docs, including Google's NotebookLM help page, xAI's Voice Agent Builder page, Anthropic's Fable redeployment note, Cursor's iOS post, Google's Gemini Spark update, and Hermes docs.
Link Map
Here is the full 01-16 map from the episode, organized as a practical builder checklist.
| # | Launch | What it does | Builder takeaway |
|---|---|---|---|
| 01 | NotebookLM Short Video Overviews | Turns complex sources into short vertical explainers that deep-dive one concept. | Useful for training, internal education, and social learning content, but review accuracy and visual sameness. |
| 02 | How to Succeed at MrBeast Production doc | An internal-style guide about how to succeed inside the MrBeast production machine. | Good source material for testing whether NotebookLM can turn dense process docs into explainers. |
| 03 | NotebookLM product home | Google's AI notebook product for summaries, audio, source-grounded chat, and video overviews. | Think of it as a source-to-learning workspace, not just a chatbot over documents. |
| 04 | June - private, local Mac AI | A privacy-first Mac app for dictation, meeting notes, chat, and on-device agents. | Local AI is becoming productized for normal operators, not only terminal users. |
| 05 | xAI Voice Agent Builder announcement | xAI's no-code voice-agent builder for Grok Voice, shown as a fast way to create agents that can call and receive calls. | Voice agents are becoming easier to build, but compliance and call permissions matter. |
| 06 | Grok Voice product page | xAI says Grok Voice Agent Builder includes a free number or bring-your-own number, 25+ languages, tool connections, guardrails, and $0.05 per minute pricing. | Test with internal calls and clear handoff rules before putting an agent in front of customers. |
| 07 | Boring Marketer brain-in-a-box agent | A demo of cloning a marketing brain into a Slack agent that the team can work with. | This is the practical future of "company knowledge": one agent per function, scoped by permissions. |
| 08 | Boring Marketing / Boring Agent | A Slack-integrated marketing agent for keyword research, AI-citation tracking, competitor analysis, and content strategy. | Marketing agents are most valuable when tied to data sources, search visibility, and repeatable reporting. |
| 09 | Open-source org-level Agent Harness | A demo of an open-source Slack agent harness where teams can bring their own model, runtime, and tools. | The open-source version of Claude Tag matters for teams that want more control over models and data flow. |
| 10 | Matt Pocock's /wizard skill | A skill that builds an interactive CLI for third-party setup and repeated workflows. | Great agent UX is often a setup wizard, not another prompt wall. |
| 11 | Arena42 / NetMind Agent Arena | A competitive benchmarking arena where autonomous AI agents go head-to-head on real-world tasks for rewards. | Agent evaluation is becoming a product category. Keep the incentives clean and task design safe. |
| 12 | Claude Fable 5 is coming back | Anthropic says Fable 5 returns globally with new cybersecurity classifiers after talks with the US government. | Frontier access can change overnight. Build fallbacks, usage limits, and routing logic. |
| 13 | Cursor for iOS | Cursor launches an iOS app to run cloud agents or control agents on your computer. | The phone becomes the review and steering layer for long-running coding work. |
| 14 | OpenClaw goes mobile | OpenClaw ships native iOS and Android apps for agents, channels, and tasks. | Mobile agent control is becoming table stakes, even if first versions are rough. |
| 15 | Gemini Spark comes to macOS | Google adds Gemini Spark to Mac with custom MCP support, app integrations, and Tasks/Keep connections. | Desktop action modes will compete with Codex, Claude Cowork, Hermes, and Cursor agents. |
| 16 | Per-session context usage breakdown | Nous Research adds a per-session breakdown so users can see what is filling the context window. | Context observability is not glamorous, but it is how you debug cost, memory, and skill bloat. |
The Pattern Across The Week
The pattern is not "AI can do more." That is too vague to be useful. The pattern is that AI is being packaged into specific workflow surfaces:
- Source-to-media learning: NotebookLM turns documents into short video explainers.
- Local/private productivity: June pushes meeting notes, dictation, chat, and agents onto the Mac.
- Voice operations: xAI lowers the setup friction for phone-call agents.
- Slack work agents: BoringMarketing and open agent harnesses move AI into team chat.
- Setup UX: Matt Pocock's /wizard skill points to repeatable installation and configuration flows.
- Agent evaluation: Arena42 turns agent performance into a competitive surface.
- Mobile control: Cursor and OpenClaw make agents steerable from phones.
- Context observability: Hermes shows where the context window is going.
That is why this week is interesting. The AI tool market is no longer only "which model is smartest?" It is also "which workflow gets packaged well enough for a real team to use?"
Short Video Learning
NotebookLM Short Video Overviews may look like a content feature, but the more useful interpretation is training infrastructure. Google says Video Overviews transform notebook sources into an engaging video, and the Short format is designed to explain key concepts in about 60 seconds. Google also warns that AI-generated videos can contain inaccuracies or audio glitches, and notes that Cinematic and Short Video Overviews currently support English and are available to eligible Ultra and Pro subscribers.
In the episode, Andrew tests it with the MrBeast production document and frames the result as a TikTok-style explainer. Corey immediately goes to education: middle schools, high schools, and training departments already fight for attention in a short-form world. Turning approved source material into short explainers could be genuinely useful.
The caveat is creative sameness. If every explainer uses the same visual rhythm, it becomes background noise fast. Use this for onboarding, internal training, and quick concept refreshers first. Treat public social channels as a second step, after human review, visual direction, and fact checking.
Agents At Work: Voice, Slack, And Marketing
The xAI Voice Agent Builder is one of the sharper business updates because it removes painful setup work. xAI describes it as a beta no-code platform for Grok Voice with telephony, knowledge retrieval, tools, guardrails, MCPs, and observability. The product page says it can deploy human-like voice agents with a free phone number or bring-your-own number option, 25+ languages, and $0.05 per minute pricing.
That does not mean every business should let an AI agent call customers tomorrow. Voice agents touch consent, call recording, local telecom rules, TCPA/CAN-SPAM-adjacent outreach risk, customer support escalation, and brand trust. But for internal demos, appointment triage, first-party inbound support, and structured phone research, this is worth testing.
BoringMarketing shows another work surface: Slack. Instead of asking a general chatbot for "SEO ideas," a Slack-based marketing agent can watch rankings, competitors, keyword data, AI-citation data, and content opportunities. That is more valuable because the agent has a job, a context, and a communication channel.
The open-source org-level Agent Harness is the complementary idea. If Claude Tag is the polished enterprise product, an open harness is the control path: bring your own model, runtime, tools, and data policies. That matters for small technical teams that want Slack agents without handing every workflow to one vendor.
Local And Private Agents
June is the quieter launch, but it points at a serious buyer need: private AI that runs on the Mac. The episode compares it to tools like Granola, but with local processing for meeting notes and dictation, plus chat and agent mode.
The strategic question is not whether every local model is better than the cloud. It is whether some workflows should stay on your machine by default:
- private meeting notes;
- drafts over sensitive internal docs;
- voice dictation in poor internet conditions;
- personal knowledge search;
- small agent routines that do not need a frontier model.
Local AI also changes resilience. If frontier access becomes gated, usage-limited, or expensive, local and routed models become the fallback layer. They may not win every hard task, but they can keep normal work moving.
Agent UX: Wizards, Arenas, Phones, And Context Bars
Matt Pocock's /wizard skill is a small signal with a large lesson. Agent setup is still too annoying. People do not want to read docs, click through dashboards, copy keys, test environment variables, and remember six commands every time. A skill that turns setup into an interactive CLI is exactly the kind of glue agents need.
Arena42 points to the evaluation side. If builders claim their agents are better, let them compete on real tasks. That is useful if the tasks are well designed, scored fairly, and safe. It is dangerous if incentives reward shortcuts, spam, or unsafe automation. Agent arenas need more than leaderboards. They need task governance.
Cursor for iOS and OpenClaw mobile are another UX lesson. Agents often take minutes, not seconds. That makes mobile review useful. Cursor says its iOS app lets users launch cloud agents, control computer-running agents from a phone, use voice input, add screenshots as context, get notifications, review diffs, and merge PRs. That is the right model: your phone is the command bridge, not the whole workstation.
Hermes context usage breakdown may sound small, but for agent operators it is gold. When skills, system prompts, memory, tools, and chat history all compete for context, you need to know what is filling the window. Otherwise you are debugging with vibes.
What Builders Should Test First
If you are a small business owner, solo builder, consultant, or technical operator, here is the sensible order:
- NotebookLM video test: pick one internal SOP or dense article and generate a Short Video Overview. Check accuracy, clarity, and whether the format actually teaches.
- Local AI test: try one private meeting-note or dictation workflow with June or your existing local stack. Compare quality, latency, and privacy comfort.
- Voice-agent test: build a sandbox voice agent for one internal scenario. Do not let it call real customers until escalation and compliance are clear.
- Slack-agent test: create one read-only marketing or research agent in Slack. Ask it for a weekly report, not autonomous execution.
- Wizard test: turn one repeated setup task into a CLI-style wizard or skill. If it saves ten minutes every time, it is worth keeping.
- Mobile coding test: use Cursor iOS or another mobile agent surface only for review, feedback, and small scoped tasks first.
- Context audit: in Hermes or any agent desktop, inspect what fills context before blaming the model.
The best tools this week are not the ones with the flashiest demos. They are the ones that reduce setup friction, respect permissions, expose what the agent is doing, and let a human steer at the right moments.
CTA: Pick one workflow surface this week: source-to-video learning, local notes, voice calls, Slack agents, mobile coding, or context observability. Turn one repeated task into a controlled system before you chase the next launch.
Sources
- The Next New Thing video: Free Claude Tag + agents battle + TikTok video maker + more
- The Next New Thing
- The Next New Thing: Links We Covered PDF, July 2, 2026
- NotebookLM Short Video Overviews X post
- Google NotebookLM Help: Generate Video Overviews
- NotebookLM product home
- How to Succeed at MrBeast Production document
- June private local Mac AI
- xAI Voice Agent Builder X post
- xAI: Introducing the Voice Agent Builder
- Grok Voice product page
- Boring Marketer brain-in-a-box X post
- Boring Marketing / Boring Agent
- Open-source org-level Agent Harness X post
- Matt Pocock /wizard skill X post
- Matt Pocock skills repository
- Arena42 / NetMind Agent Arena
- Anthropic Fable 5 returning X post
- Anthropic: Redeploying Fable 5
- Cursor for iOS X post
- Cursor: Build from anywhere with Cursor for iOS
- OpenClaw mobile X post
- Google Gemini Spark on macOS X post
- Google: Gemini Spark updates for macOS, connected apps, and more
- Hermes per-session context usage X post
- Hermes Agent documentation