Google's agent computers lead today's AI brief

Google announces dedicated computers for agents · 22 October

Agent Compute Engine is designed for agents that operate browsers, execute code and retain a workspace. Google says the external preview will arrive in the coming months.

• Each small virtual machine gets a separate boundary and identity. Service credentials can be injected outside that machine.
• Workspaces can retain state, while idle sessions can pause and resume.
• Agent Substrate is a separate open-source Kubernetes runtime mentioned in the announcement. Do not confuse it with availability of the managed service.

For a deployment plan, check launch regions, pricing and release status. Those details are not established here.

Anthropic transfers Agent Skills to the AAIF · 22 October

Anthropic is transferring the Agent Skills standard to the Linux Foundation’s Agentic AI Foundation. The news changes its governance, rather than launching another model.

• A skill packages a specific procedure in a SKILL.md file, with optional scripts and references.
• For example, an enquiry-triage skill could read a fictional lead, classify the request and draft a response. This is an illustration, not a tested customer workflow.
• The specification repository below explains the format. It is not a ready-made business skill.

1. Read the chosen skill and any bundled scripts before installing it.
2. Confirm assistant support, dependencies, licence and model costs. Use dummy inputs first.

The same instructions may travel between compatible agents; their tools, permissions and output can still differ.

1Password announces approved AI sign-ins · 22 October

An agent can request a login without receiving the password. The new platform fills the destination form after your approval.

• Muse and Grok Bot integrations are rolling out; ChatGPT Work, dots and Duet are coming soon.
• Eligible paid Individual/Family accounts on 1Password.com have no extra platform charge; .eu/.ca accounts are unsupported.

Check partner support. Hidden credentials do not limit everything an agent can do after login.

Uber expands its workplace-agent rollout · 22 October

OpenAI says Uber is expanding ChatGPT Work from a 1,000-person pilot to more than 30,000 employees. This is an enterprise deployment, not a new consumer offer.

• The rollout includes approved integrations, an internal skills library and employee training.
• Examples include preparing meeting briefs, presentations and analysis using connected company tools.
• The announcement does not provide measured productivity gains; the companies say outcomes will be shared later.

The practical lesson is to plan approved data access and staff training alongside the software.

Pictory 3 turns source material into a video draft · 22 October

The new video agent accepts material such as a brief, document or URL. It builds a script and scenes for review, rather than making the first draft your final approval.

• Setup choices cover voice, visual sources, format, layout and brand kit.
• Users can revise the draft through chat or the editor before generating an export.
• New users enter the new experience as it rolls out; older projects keep their existing editor. Pricing and export limits need checking.

1. Start with an approved feature brief and assets you can use.
2. Review names, numbers, product claims and visual rights before export.

This is a vendor-described workflow, not a tool test by Infyny Labs.

How to use the five skills below

These are reusable instructions for a coding assistant, not five standalone apps. They are established picks from two GitHub repositories, selected for useful business tasks; this is not a claim that they launched today or went viral.

• Code and instructions use Apache2.0 or MIT licences. Keep the applicable licence notices. Model usage and optional external tools may still cost money.
• Nothing below was installed or run for this news edition. Browser and command access depend on your assistant and environment.

1. Read the linked SKILL.md and licence first. Download the complete named folder, preserving its scripts and references.
2. For Codex, place the folder under your project’s .agents/skills/<skill-name>/ and invoke it with $<skill-name>, followed by your task. An alternative is to ask the built-in skill-installer to install the exact linked folder.
3. For Claude Code, place it under .claude/skills/<skill-name>/ and request /<skill-name> with your task. The Anthropic README also documents its example-skills plugin.
4. Provide the inputs listed below. Add required tools, then try dummy data or a local test project. Review the result before using it with customers.

A link alone does not install a skill or grant it access. Start with one job and a result you can check.

Build an interface from a real brief · frontend-design

Turn your brand brief into a working website interface.

• Input: Audience, real page content, brand colours and existing project.
• Output: Interface code plus a design direction; inspect the result on mobile and desktop.
• Needs: Coding agent with project file access and the project's frontend runtime; browser/screenshots useful for review. No special fixed runtime declared in this skill.

1. Install the complete frontend-design folder using the shared setup above, then supply the listed inputs.
2. Example prompt: Use frontend-design to build a landing page for my invoicing product. Keep the supplied burgundy/yellow brand, explain its real features, and make the enquiry form usable on mobile.

Licence: Apache-2.0. Review the output in your own environment; installation alone does not guarantee the result.

Test the enquiry form in a browser · webapp-testing

Test your website's forms and buttons in a browser.

• Input: A local test app, expected behaviour and permitted test data.
• Output: Python Playwright tests, screenshots and observed browser logs. It does not prove every defect is absent.
• Needs: Python, Playwright and its browser; running local app or complete scripts/with_server.py helper folder; app-specific startup dependencies.

1. Install the complete webapp-testing folder using the shared setup above, then supply the listed inputs.
2. Example prompt: Use webapp-testing on my local staging app. Check the enquiry form's required fields and successful submission with dummy data; record screenshots and console errors. Do not submit to production.

Licence: Apache-2.0. Review the output in your own environment; installation alone does not guarantee the result.

Find and prioritise SEO issues · seo-audit

Find technical and content problems on your website.

• Input: Website URL, business goal, target audience and optional Search Console/crawl exports.
• Output: Prioritised findings with evidence, impact and recommended fixes. No ranking guarantee.
• Needs: Agent web/browser or supplied exports; optional Search Console/PageSpeed/crawler data. Optional paid SEO tools cost separately. Text fetch alone can miss JavaScript-injected schema.

1. Install the complete seo-audit folder using the shared setup above, then supply the listed inputs.
2. Example prompt: Use seo-audit to inspect my public service pages. Prioritise indexability, canonical URLs, internal links and titles. Cite evidence, distinguish untested items, and do not change the site.

Licence: MIT. Review the output in your own environment; installation alone does not guarantee the result.

Make your homepage offer clear · copywriting

Rewrite your homepage so buyers understand the offer.

• Input: Current page, real product facts, audience, desired action and brand voice.
• Output: Revised page copy with headline/CTA options and rationale; unsupported claims remain flagged.
• Needs: No fixed runtime required for drafting. AI agent file access to the full skill and references. Optional integrations/credentials, external services and separate authorization are needed for implementation; never imply automatic sending.

1. Install the complete copywriting folder using the shared setup above, then supply the listed inputs.
2. Example prompt: Use copywriting to rewrite this homepage for small business owners. Keep verified services and our voice, explain the offer plainly and use one enquiry action. Do not invent results.

Licence: MIT. Review the output in your own environment; installation alone does not guarantee the result.

Draft follow-up emails for interested leads · emails

Plan useful follow-up emails for interested leads.

• Input: Consent-based audience, entry trigger, offer, goal and existing email schedule.
• Output: Draft sequence with subject, preview, body, CTA, timing and exit conditions. Installing it does not send emails.
• Needs: No fixed runtime required for drafting. AI agent file access to the full skill and references. Optional integrations/credentials, external services and separate authorization are needed for implementation; never imply automatic sending.

1. Install the complete emails folder using the shared setup above, then supply the listed inputs.
2. Example prompt: Use emails to draft three follow-up messages for people who requested our demo. Explain the product, answer one objection and offer a booking link. Draft only; do not send.

Licence: MIT. This produces draft emails; connecting a sender and authorising delivery are separate.

Model watch: DNA and language in Carbon-X · released 22 October

Carbon-X combines DNA sequences and natural-language questions in a downloadable research model. Its release history says 22 October; later coverage should not become a new launch date.

• The model card lists Apache 2.0 licensing and a 14.8B-parameter model. Compute costs still apply.
• It has not been clinically validated and must not guide medical decisions.

Read the model card and hardware requirements before planning a research trial.

Developer catch-up: enterprise agent toolkit · 21 October

NVIDIA describes an expanded stack for connecting agents to company knowledge and running them with access boundaries.

• The announcement includes data-retrieval components, the OpenShell runtime and agent blueprints.
• Component licences, dependencies and hosting needs must be checked individually. A toolkit is not the same thing as a SKILL.md package.

Treat the announcement as a set of engineering components, not an already integrated business application.

Platform catch-up: OpenAI and Google Cloud · 21 October

The companies announced plans to offer OpenAI models and Codex workflows through Google Cloud.

• Model Garden access and Marketplace billing are part of the announced route.
• General availability is planned in the coming weeks. Check models, regions and pricing when access opens.

An announced partnership is not confirmation that every customer can enable it today.

Earlier Indian security context: CERT-In · 15 October

This is older context, not a new 23 October alert. CERT-In described evaluation agents exceeding their permissions and separate deliberate misuse of agentic tools.

• The advisory discusses exposed configurations, repositories and overly permissive access.
• It does not name every provider or establish a 100-company victim count.

Review the accounts and systems an agent can reach; approving a login is only one control.

Research catch-up: a model for longer investigations · report 20 October

The Sage-A1.5 report describes a research agent that delegates work and checks evidence during extended tasks.

• Its downloadable model card lists Apache 2.0; tools and an agent runtime still require setup.
• The evaluation results come from the authors. They do not prove that it beats every larger model on general work.

Test it on a bounded question with references you can verify.

Product catch-up: Claude Duet · 21 October

Duet is a research-preview background agent for connected apps. The initial waitlist targets US Pro and Max subscribers.

• Background work shares Claude usage allowances and sensitive actions require approval.
• The announcement does not establish general India or Middle East access. The new 1Password integration is still coming soon.

Confirm eligibility and enabled integrations before making this part of a client workflow.