Daily AI · September 23, 2026
Useful AI Daily - September 23, 2026
AI is moving from a reply box into code review, local shells, glasses and business dashboards. The useful move today is not to hand it everything—it is to set a task, a boundary and a review point that a human can understand.
September 23's releases give ordinary users, creators and builders a practical pattern: use AI to make a recommendation or a draft first, run local agents with the least access possible, and keep high-impact decisions easy to inspect and undo.
The Short Version
- Automatic code review needs an explicit effort level and a human merge decision.
- A local agent should start with denied folders, limited network access and no credentials.
- Hands-free navigation and assistants can be useful without becoming an always-on recorder.
- Wearable AI needs clear capture indicators and a bystander-conscious habit.
- Small-business automation is safest when it proposes before it acts.
5 Updates Worth Your Time
Make automated code review a conscious setting
- What changed
- GitHub expanded Copilot code-review controls with a dedicated personal settings page, configurable automatic reviews and selectable review effort, plus an enterprise-level default for organization repositories.
- Why it matters
- A lightweight review can help early feedback, while a more deliberate pass may suit risky changes. Neither should substitute for the human decision to merge.
- Try, watch, or skip?
- Enable one automatic review mode on a low-risk repository, compare Lite and Balanced feedback on three pull requests, and keep a required human review for meaningful changes.
Give local coding agents less to touch
- What changed
- GitHub released public-preview local sandboxing for new Copilot app sessions, with per-project controls for filesystem access, network access and selected Git or GitHub CLI credentials.
- Why it matters
- A coding assistant that makes an unexpected command has a smaller blast radius when it cannot reach unrelated files, the open internet or your credentials.
- Try, watch, or skip?
- Create a test project with one allowed folder, no credentials and outbound network off. Run a harmless task and expand access only when the work requires it.
Use heads-up AI features without giving up attention
- What changed
- Meta announced hands-free cycling and transit navigation, a video-call hologram feature, an Explore app and additional calendar, notification and capture capabilities for Meta Ray-Ban Display.
- Why it matters
- For an ordinary user, a screen that keeps a phone in a pocket can be useful, but notifications, camera use and location prompts are easy to overlook when the interface is in front of your eyes.
- Try, watch, or skip?
- Start with navigation only. Turn off mirrored notifications you do not need, verify capture indicators and avoid using visual AI features where people have a reasonable expectation of privacy.
Choose wearable AI for a clear job, not a vague upgrade
- What changed
- Meta introduced Ray-Ban Meta Audio glasses and described hands-free calls, media and assistant access, alongside broader AI-glasses options and future Muse integration.
- Why it matters
- New hardware can make translation, reminders and calls more convenient, but it also creates a new place for microphones, account connections and voice history to accumulate.
- Try, watch, or skip?
- Decide on one use case before buying or connecting accounts. Review voice history, device-sharing, capture controls and how quickly you can disable the assistant.
Let AI prepare the business decision before it makes it
- What changed
- Amazon's Seller Assistant materials describe AI-powered recommendations and, with seller permission, actions across operational work such as account health, inventory and listings.
- Why it matters
- Indie operators can save time when an assistant assembles the context, but a recommendation affecting stock, compliance or customer-facing information still needs a visible review step.
- Try, watch, or skip?
- Ask for a weekly exceptions list first. Verify one recommendation against source data and approve only a reversible action before allowing wider automation.
Tool Worth Trying Today
A least-privilege local-agent setup
Create a small disposable project, deny access to everything outside it, switch off credentials and network access, then see whether an agent can still complete a narrow coding task.
Best for: Developers and technical indie builders trying a local coding agent without exposing their whole workstation.
Watch out: Sandboxing is not a substitute for code review. Keep secrets out of the test project and inspect every generated change before running it.
Privacy / Cost Watch
- Do not give a new local agent access to personal, customer, legal, unreleased or private photo/document data until its permissions and retention settings are clear.
- Use dedicated or disposable credentials for a sandbox trial; never make a broad personal token the default convenience option.
- Wearable microphones, cameras, notifications and location features affect both the wearer and people nearby. Learn the indicators and turn off unused connections.
- Business recommendations may reflect stale or incomplete data. Verify inventory, price and compliance information in the source system before acting.
One Practical Workflow
Run a 25-minute least-privilege AI trial
- Choose one small task that ends in a draft, recommendation or local code change.
- List the files, network destinations, accounts and personal data the task truly needs.
- Start with every other path denied and use public or synthetic material for the first run.
- Review the output with a human before it is merged, sent, published or applied to a business record.
- Record the useful permission, the unnecessary permission and the rollback step for the next trial.
Builder Note
Build interfaces that make the narrow path obvious: one folder, one account, one proposed action and one clear approve button. That is easier to trust than an agent that asks for the whole computer on day one.
Ignore For Now
Full-access agent demos
A demo that succeeds only after broad filesystem, network, credential or account access is not a deployment plan. Skip the spectacle and test whether the job still works with a narrow, reviewable scope.
Bottom Line
September 23 is a day for constrained trials. New code-review controls, local sandboxes, wearable assistants and seller tools are most useful when they begin with the minimum access and a human approval point.
Sources
- GitHub Changelog: More ways to request and configure Copilot code reviews
- GitHub Changelog: Local sandboxing in the GitHub Copilot app
- Meta: New Features for Meta Ray-Ban Display
- Meta: Introducing Ray-Ban Meta Audio and More AI Glasses Styles
- Amazon Selling Partners: Why Amazon's new AI seller tools feel like a turning point